Robby T. Tan

Chenxin Xu, Robby T. Tan, Yuhong Tan et al. · cambridge

Learning to predict agent motions with relationship reasoning is important for many applications. In motion prediction tasks, maintaining motion equivariance under Euclidean geometric transformations and invariance of agent interaction is a critical and fundamental principle. However, such equivariance and invariance properties are overlooked by most existing methods. To fill this gap, we propose EqMotion, an efficient equivariant motion prediction model with invariant interaction reasoning. To achieve motion equivariance, we propose an equivariant geometric feature learning module to learn a Euclidean transformable feature through dedicated designs of equivariant operations. To reason agent's interactions, we propose an invariant interaction reasoning module to achieve a more stable interaction modeling. To further promote more comprehensive motion features, we propose an invariant pattern feature learning module to learn an invariant pattern feature, which cooperates with the equivariant geometric feature to enhance network expressiveness. We conduct experiments for the proposed model on four distinct scenarios: particle dynamics, molecule dynamics, human skeleton motion prediction and pedestrian trajectory prediction. Experimental results show that our method is not only generally applicable, but also achieves state-of-the-art prediction performances on all the four tasks, improving by 24.0/30.1/8.6/9.2%. Code is available at https://github.com/MediaBrain-SJTU/EqMotion.

Yeying Jin, Aashish Sharma, Robby T. Tan

Shadow removal from a single image is generally still an open problem. Most existing learning-based methods use supervised learning and require a large number of paired images (shadow and corresponding non-shadow images) for training. A recent unsupervised method, Mask-ShadowGAN~\cite{Hu19}, addresses this limitation. However, it requires a binary mask to represent shadow regions, making it inapplicable to soft shadows. To address the problem, in this paper, we propose an unsupervised domain-classifier guided shadow removal network, DC-ShadowNet. Specifically, we propose to integrate a shadow/shadow-free domain classifier into a generator and its discriminator, enabling them to focus on shadow regions. To train our network, we introduce novel losses based on physics-based shadow-free chromaticity, shadow-robust perceptual features, and boundary smoothness. Moreover, we show that our unsupervised network can be used for test-time training that further improves the results. Our experiments show that all these novel components allow our method to handle soft shadows, and also to perform better on hard shadows both quantitatively and qualitatively than the existing state-of-the-art shadow removal methods. Our code is available at: \url{https://github.com/jinyeying/DC-ShadowNet-Hard-and-Soft-Shadow-Removal}.

Chenxin Xu, Robby T. Tan, Yuhong Tan et al.

Exploring spatial-temporal dependencies from observed motions is one of the core challenges of human motion prediction. Previous methods mainly focus on dedicated network structures to model the spatial and temporal dependencies. This paper considers a new direction by introducing a model learning framework with auxiliary tasks. In our auxiliary tasks, partial body joints' coordinates are corrupted by either masking or adding noise and the goal is to recover corrupted coordinates depending on the rest coordinates. To work with auxiliary tasks, we propose a novel auxiliary-adapted transformer, which can handle incomplete, corrupted motion data and achieve coordinate recovery via capturing spatial-temporal dependencies. Through auxiliary tasks, the auxiliary-adapted transformer is promoted to capture more comprehensive spatial-temporal dependencies among body joints' coordinates, leading to better feature learning. Extensive experimental results have shown that our method outperforms state-of-the-art methods by remarkable margins of 7.2%, 3.7%, and 9.4% in terms of 3D mean per joint position error (MPJPE) on the Human3.6M, CMU Mocap, and 3DPW datasets, respectively. We also demonstrate that our method is more robust under data missing cases and noisy data cases. Code is available at https://github.com/MediaBrain-SJTU/AuxFormer.

Yeying Jin, Beibei Lin, Wending Yan et al.

Visibility in hazy nighttime scenes is frequently reduced by multiple factors, including low light, intense glow, light scattering, and the presence of multicolored light sources. Existing nighttime dehazing methods often struggle with handling glow or low-light conditions, resulting in either excessively dark visuals or unsuppressed glow outputs. In this paper, we enhance the visibility from a single nighttime haze image by suppressing glow and enhancing low-light regions. To handle glow effects, our framework learns from the rendered glow pairs. Specifically, a light source aware network is proposed to detect light sources of night images, followed by the APSF (Atmospheric Point Spread Function)-guided glow rendering. Our framework is then trained on the rendered images, resulting in glow suppression. Moreover, we utilize gradient-adaptive convolution, to capture edges and textures in hazy scenes. By leveraging extracted edges and textures, we enhance the contrast of the scene without losing important structural details. To boost low-light intensity, our network learns an attention map, then adjusted by gamma correction. This attention has high values on low-light regions and low values on haze and glow regions. Extensive evaluation on real nighttime haze images, demonstrates the effectiveness of our method. Our experiments demonstrate that our method achieves a PSNR of 30.38dB, outperforming state-of-the-art methods by 13% on GTA5 nighttime haze dataset. Our data and code is available at https://github.com/jinyeying/nighttime_dehaze.

Yeying Jin, Wei Ye, Wenhan Yang et al.

Removing soft and self shadows that lack clear boundaries from a single image is still challenging. Self shadows are shadows that are cast on the object itself. Most existing methods rely on binary shadow masks, without considering the ambiguous boundaries of soft and self shadows. In this paper, we present DeS3, a method that removes hard, soft and self shadows based on adaptive attention and ViT similarity. Our novel ViT similarity loss utilizes features extracted from a pre-trained Vision Transformer. This loss helps guide the reverse sampling towards recovering scene structures. Our adaptive attention is able to differentiate shadow regions from the underlying objects, as well as shadow regions from the object casting the shadow. This capability enables DeS3 to better recover the structures of objects even when they are partially occluded by shadows. Different from existing methods that rely on constraints during the training phase, we incorporate the ViT similarity during the sampling stage. Our method outperforms state-of-the-art methods on the SRD, AISTD, LRSS, USR and UIUC datasets, removing hard, soft, and self shadows robustly. Specifically, our method outperforms the SOTA method by 16\% of the RMSE of the whole image on the LRSS dataset. Our data and code is available at: \url{https://github.com/jinyeying/DeS3_Deshadow}

Yeying Jin, Ruoteng Li, Wenhan Yang et al.

Estimating the reflectance layer from a single image is a challenging task. It becomes more challenging when the input image contains shadows or specular highlights, which often render an inaccurate estimate of the reflectance layer. Therefore, we propose a two-stage learning method, including reflectance guidance and a Shadow/Specular-Aware (S-Aware) network to tackle the problem. In the first stage, an initial reflectance layer free from shadows and specularities is obtained with the constraint of novel losses that are guided by prior-based shadow-free and specular-free images. To further enforce the reflectance layer to be independent of shadows and specularities in the second-stage refinement, we introduce an S-Aware network that distinguishes the reflectance image from the input image. Our network employs a classifier to categorize shadow/shadow-free, specular/specular-free classes, enabling the activation features to function as attention maps that focus on shadow/specular regions. Our quantitative and qualitative evaluations show that our method outperforms the state-of-the-art methods in the reflectance layer estimation that is free from shadows and specularities. Code is at: \url{https://github.com/jinyeying/S-Aware-network}.

Yu Cheng, Bo Wang, Robby T. Tan

Monocular 3D human pose estimation has made progress in recent years. Most of the methods focus on single persons, which estimate the poses in the person-centric coordinates, i.e., the coordinates based on the center of the target person. Hence, these methods are inapplicable for multi-person 3D pose estimation, where the absolute coordinates (e.g., the camera coordinates) are required. Moreover, multi-person pose estimation is more challenging than single pose estimation, due to inter-person occlusion and close human interactions. Existing top-down multi-person methods rely on human detection (i.e., top-down approach), and thus suffer from the detection errors and cannot produce reliable pose estimation in multi-person scenes. Meanwhile, existing bottom-up methods that do not use human detection are not affected by detection errors, but since they process all persons in a scene at once, they are prone to errors, particularly for persons in small scales. To address all these challenges, we propose the integration of top-down and bottom-up approaches to exploit their strengths. Our top-down network estimates human joints from all persons instead of one in an image patch, making it robust to possible erroneous bounding boxes. Our bottom-up network incorporates human-detection based normalized heatmaps, allowing the network to be more robust in handling scale variations. Finally, the estimated 3D poses from the top-down and bottom-up networks are fed into our integration network for final 3D poses. To address the common gaps between training and testing data, we do optimization during the test time, by refining the estimated 3D human poses using high-order temporal constraint, re-projection loss, and bone length regularizations. Our evaluations demonstrate the effectiveness of the proposed method. Code and models are available: https://github.com/3dpose/3D-Multi-Person-Pose.

Yihao Ai, Yifei Qi, Bo Wang et al.

Existing 2D human pose estimation research predominantly concentrates on well-lit scenarios, with limited exploration of poor lighting conditions, which are a prevalent aspect of daily life. Recent studies on low-light pose estimation require the use of paired well-lit and low-light images with ground truths for training, which are impractical due to the inherent challenges associated with annotation on low-light images. To this end, we introduce a novel approach that eliminates the need for low-light ground truths. Our primary novelty lies in leveraging two complementary-teacher networks to generate more reliable pseudo labels, enabling our model achieves competitive performance on extremely low-light images without the need for training with low-light ground truths. Our framework consists of two stages. In the first stage, our model is trained on well-lit data with low-light augmentations. In the second stage, we propose a dual-teacher framework to utilize the unlabeled low-light data, where a center-based main teacher produces the pseudo labels for relatively visible cases, while a keypoints-based complementary teacher focuses on producing the pseudo labels for the missed persons of the main teacher. With the pseudo labels from both teachers, we propose a person-specific low-light augmentation to challenge a student model in training to outperform the teachers. Experimental results on real low-light dataset (ExLPose-OCN) show, our method achieves 6.8% (2.4 AP) improvement over the state-of-the-art (SOTA) method, despite no low-light ground-truth data is used in our approach, in contrast to the SOTA method. Our code will be available at:https://github.com/ayh015-dev/DA-LLPose.

Xin Yang, Michael Bi Mi, Yuan Yuan et al.

Most existing domain adaptation (DA) methods align the features based on the domain feature distributions and ignore aspects related to fog, background and target objects, rendering suboptimal performance. In our DA framework, we retain the depth and background information during the domain feature alignment. A consistency loss between the generated depth and fog transmission map is introduced to strengthen the retention of the depth information in the aligned features. To address false object features potentially generated during the DA process, we propose an encoder-decoder framework to reconstruct the fog-free background image. This reconstruction loss also reinforces the encoder, i.e., our DA backbone, to minimize false object features.Moreover, we involve our target data in training both our DA module and our detection module in a semi-supervised manner, so that our detection module is also exposed to the unlabeled target data, the type of data used in the testing stage. Using these ideas, our method significantly outperforms the state-of-the-art method (47.6 mAP against the 44.3 mAP on the Foggy Cityscapes dataset), and obtains the best performance on multiple real-image public datasets. Code is available at: https://github.com/VIML-CVDL/Object-Detection-in-Foggy-Scenes

Xin Li, Yeying Jin, Suhang Yao et al.

This paper presents an overview of the NTIRE 2026 Second Challenge on Day and Night Raindrop Removal for Dual-Focused Images. Building upon the success of the first edition, this challenge attracted a wide range of impressive solutions, all developed and evaluated on our real-world Raindrop Clarity dataset~\cite{jin2024raindrop}. For this edition, we adjust the dataset with 14,139 images for training, 407 images for validation, and 593 images for testing. The primary goal of this challenge is to establish a strong and practical benchmark for the removal of raindrops under various illumination and focus conditions. In total, 168 teams have registered for the competition, and 17 teams submitted valid final solutions and fact sheets for the testing phase. The submitted methods achieved strong performance on the Raindrop Clarity dataset, demonstrating the growing progress in this challenging task.

Yeying Jin, Wenhan Yang, Robby T. Tan

Night images suffer not only from low light, but also from uneven distributions of light. Most existing night visibility enhancement methods focus mainly on enhancing low-light regions. This inevitably leads to over enhancement and saturation in bright regions, such as those regions affected by light effects (glare, floodlight, etc). To address this problem, we need to suppress the light effects in bright regions while, at the same time, boosting the intensity of dark regions. With this idea in mind, we introduce an unsupervised method that integrates a layer decomposition network and a light-effects suppression network. Given a single night image as input, our decomposition network learns to decompose shading, reflectance and light-effects layers, guided by unsupervised layer-specific prior losses. Our light-effects suppression network further suppresses the light effects and, at the same time, enhances the illumination in dark regions. This light-effects suppression network exploits the estimated light-effects layer as the guidance to focus on the light-effects regions. To recover the background details and reduce hallucination/artefacts, we propose structure and high-frequency consistency losses. Our quantitative and qualitative evaluations on real images show that our method outperforms state-of-the-art methods in suppressing night light effects and boosting the intensity of dark regions.

Jiadong Wang, Xinyuan Qian, Malu Zhang et al.

Talking face generation, also known as speech-to-lip generation, reconstructs facial motions concerning lips given coherent speech input. The previous studies revealed the importance of lip-speech synchronization and visual quality. Despite much progress, they hardly focus on the content of lip movements i.e., the visual intelligibility of the spoken words, which is an important aspect of generation quality. To address the problem, we propose using a lip-reading expert to improve the intelligibility of the generated lip regions by penalizing the incorrect generation results. Moreover, to compensate for data scarcity, we train the lip-reading expert in an audio-visual self-supervised manner. With a lip-reading expert, we propose a novel contrastive learning to enhance lip-speech synchronization, and a transformer to encode audio synchronically with video, while considering global temporal dependency of audio. For evaluation, we propose a new strategy with two different lip-reading experts to measure intelligibility of the generated videos. Rigorous experiments show that our proposal is superior to other State-of-the-art (SOTA) methods, such as Wav2Lip, in reading intelligibility i.e., over 38% Word Error Rate (WER) on LRS2 dataset and 27.8% accuracy on LRW dataset. We also achieve the SOTA performance in lip-speech synchronization and comparable performances in visual quality.

Yeying Jin, Wending Yan, Wenhan Yang et al.

Few existing image defogging or dehazing methods consider dense and non-uniform particle distributions, which usually happen in smoke, dust and fog. Dealing with these dense and/or non-uniform distributions can be intractable, since fog's attenuation and airlight (or veiling effect) significantly weaken the background scene information in the input image. To address this problem, we introduce a structure-representation network with uncertainty feedback learning. Specifically, we extract the feature representations from a pre-trained Vision Transformer (DINO-ViT) module to recover the background information. To guide our network to focus on non-uniform fog areas, and then remove the fog accordingly, we introduce the uncertainty feedback learning, which produces the uncertainty maps, that have higher uncertainty in denser fog regions, and can be regarded as an attention map that represents fog's density and uneven distribution. Based on the uncertainty map, our feedback network refines our defogged output iteratively. Moreover, to handle the intractability of estimating the atmospheric light colors, we exploit the grayscale version of our input image, since it is less affected by varying light colors that are possibly present in the input image. The experimental results demonstrate the effectiveness of our method both quantitatively and qualitatively compared to the state-of-the-art methods in handling dense and non-uniform fog or smoke.

Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli et al.

Object detection at night is a challenging problem due to the absence of night image annotations. Despite several domain adaptation methods, achieving high-precision results remains an issue. False-positive error propagation is still observed in methods using the well-established student-teacher framework, particularly for small-scale and low-light objects. This paper proposes a two-phase consistency unsupervised domain adaptation network, 2PCNet, to address these issues. The network employs high-confidence bounding-box predictions from the teacher in the first phase and appends them to the student's region proposals for the teacher to re-evaluate in the second phase, resulting in a combination of high and low confidence pseudo-labels. The night images and pseudo-labels are scaled-down before being used as input to the student, providing stronger small-scale pseudo-labels. To address errors that arise from low-light regions and other night-related attributes in images, we propose a night-specific augmentation pipeline called NightAug. This pipeline involves applying random augmentations, such as glare, blur, and noise, to daytime images. Experiments on publicly available datasets demonstrate that our method achieves superior results to state-of-the-art methods by 20\%, and to supervised models trained directly on the target data.

Yu Cheng, Yihao Ai, Bo Wang et al.

In multi-person 2D pose estimation, the bottom-up methods simultaneously predict poses for all persons, and unlike the top-down methods, do not rely on human detection. However, the SOTA bottom-up methods' accuracy is still inferior compared to the existing top-down methods. This is due to the predicted human poses being regressed based on the inconsistent human bounding box center and the lack of human-scale normalization, leading to the predicted human poses being inaccurate and small-scale persons being missed. To push the envelope of the bottom-up pose estimation, we firstly propose multi-scale training to enhance the network to handle scale variation with single-scale testing, particularly for small-scale persons. Secondly, we introduce dual anatomical centers (i.e., head and body), where we can predict the human poses more accurately and reliably, especially for small-scale persons. Moreover, existing bottom-up methods use multi-scale testing to boost the accuracy of pose estimation at the price of multiple additional forward passes, which weakens the efficiency of bottom-up methods, the core strength compared to top-down methods. By contrast, our multi-scale training enables the model to predict high-quality poses in a single forward pass (i.e., single-scale testing). Our method achieves 38.4\% improvement on bounding box precision and 39.1\% improvement on bounding box recall over the state of the art (SOTA) on the challenging small-scale persons subset of COCO. For the human pose AP evaluation, we achieve a new SOTA (71.0 AP) on the COCO test-dev set with the single-scale testing. We also achieve the top performance (40.3 AP) on OCHuman dataset in cross-dataset evaluation.

Wending Yan, Lu Xu, Wenhan Yang et al.

Existing adherent raindrop removal methods focus on the detection of the raindrop locations, and then use inpainting techniques or generative networks to recover the background behind raindrops. Yet, as adherent raindrops are diverse in sizes and appearances, the detection is challenging for both single image and video. Moreover, unlike rain streaks, adherent raindrops tend to cover the same area in several frames. Addressing these problems, our method employs a two-stage video-based raindrop removal method. The first stage is the single image module, which generates initial clean results. The second stage is the multiple frame module, which further refines the initial results using temporal constraints, namely, by utilizing multiple input frames in our process and applying temporal consistency between adjacent output frames. Our single image module employs a raindrop removal network to generate initial raindrop removal results, and create a mask representing the differences between the input and initial output. Once the masks and initial results for consecutive frames are obtained, our multiple-frame module aligns the frames in both the image and feature levels and then obtains the clean background. Our method initially employs optical flow to align the frames, and then utilizes deformable convolution layers further to achieve feature-level frame alignment. To remove small raindrops and recover correct backgrounds, a target frame is predicted from adjacent frames. A series of unsupervised losses are proposed so that our second stage, which is the video raindrop removal module, can self-learn from video data without ground truths. Experimental results on real videos demonstrate the state-of-art performance of our method both quantitatively and qualitatively.

Shuwei Li, Lei Tan, Robby T. Tan

Color constancy aims to keep object colors consistent under varying illumination. Cross-camera generalization in color constancy remains challenging because learning-based models often overfit to the color response characteristics of the training camera, resulting in degraded performance on images captured by other cameras. We propose VLM-CC, a feedback-guided framework that formulates color constancy as an iterative refinement process. Instead of directly estimating the illuminant from raw input, VLM-CC performs iterative correction driven by vision-language model (VLM)-based evaluation. At each iteration, the image is white-balanced using the current estimate and converted to pseudo-sRGB. A lightweight LoRA-tuned VLM then assesses the corrected image, identifying the dominant residual color cast and providing qualitative feedback. This feedback is mapped to a residual illumination direction (red, green, or blue) and used to update the illuminant estimate until convergence. Our key idea is to reframe color constancy as an iterative perceptual feedback problem, leveraging VLM evaluation instead of direct RGB regression. By replacing direct RGB estimation with VLM-guided perceptual feedback, VLM-CC achieves state-of-the-art robustness in cross-camera color constancy across multiple datasets. Code will be available at https://github.com/NothingIknow/VLM-CC.

Shuwei Li, Jikai Wang, Michael S. Brown et al.

The assumption of a uniform light color distribution is no longer applicable in scenes that have multiple light colors. Most color constancy methods are designed to deal with a single light color, and thus are erroneous when applied to multiple light colors. The spatial variability in multiple light colors causes the color constancy problem to be more challenging and requires the extraction of local surface/light information. Motivated by this, we introduce a multi-task learning method to discount multiple light colors in a single input image. To have better cues of the local surface/light colors under multiple light color conditions, we design a novel multi-task learning framework. Our framework includes auxiliary tasks of achromatic-pixel detection and surface-color similarity prediction, providing better cues for local light and surface colors, respectively. Moreover, to ensure that our model maintains the constancy of surface colors regardless of the variations of light colors, a novel local surface color feature preservation scheme is developed. We demonstrate that our model achieves 47.1% improvement (from 4.69 mean angular error to 2.48) compared to a state-of-the-art multi-illuminant color constancy method on a multi-illuminant dataset (LSMI).

Haopeng Chen, Yihao Ai, Kabeen Kim et al.

Low-visibility scenarios, such as low-light conditions, pose significant challenges to human pose estimation due to the scarcity of annotated low-light datasets and the loss of visual information under poor illumination. Recent domain adaptation techniques attempt to utilize well-lit labels by augmenting well-lit images to mimic low-light conditions. But handcrafted augmentations oversimplify noise patterns, while learning-based methods often fail to preserve high-frequency low-light characteristics, producing unrealistic images that lead pose models to generalize poorly to real low-light scenes. Moreover, recent pose estimators rely on image cues through image-to-keypoint cross-attention, but these cues become unreliable under low-light conditions. To address these issues, we propose Unsupervised Domain Adaptation for Pose Estimation (UDAPose), a novel framework that synthesizes low-light images and dynamically fuses visual cues with pose priors for improved pose estimation. Specifically, our synthesis method incorporates a Direct-Current-based High-Pass Filter (DHF) and a Low-light Characteristics Injection Module (LCIM) to inject high-frequency details from input low-light images, overcoming rigidity or the detail loss in existing approaches. Furthermore, we introduce a Dynamic Control of Attention (DCA) module that adaptively balances image cues with learned pose priors in the Transformer architecture. Experiments show that UDAPose outperforms state-of-the-art methods, with notable AP gains of 10.1 (56.4%) on the ExLPose-test hard set (LL-H) and 7.4 (31.4%) in cross-dataset validation on EHPT-XC. Code: https://github.com/Vision-and-Multimodal-Intelligence-Lab/UDAPose

Aishik Nagar, Yutong Liu, Andy T. Liu et al.

Medical abstractive summarization faces the challenge of balancing faithfulness and informativeness. Current methods often sacrifice key information for faithfulness or introduce confabulations when prioritizing informativeness. While recent advancements in techniques like in-context learning (ICL) and fine-tuning have improved medical summarization, they often overlook crucial aspects such as faithfulness and informativeness without considering advanced methods like model reasoning and self-improvement. Moreover, the field lacks a unified benchmark, hindering systematic evaluation due to varied metrics and datasets. This paper addresses these gaps by presenting a comprehensive benchmark of six advanced abstractive summarization methods across three diverse datasets using five standardized metrics. Building on these findings, we propose uMedSum, a modular hybrid summarization framework that introduces novel approaches for sequential confabulation removal followed by key missing information addition, ensuring both faithfulness and informativeness. Our work improves upon previous GPT-4-based state-of-the-art (SOTA) medical summarization methods, significantly outperforming them in both quantitative metrics and qualitative domain expert evaluations. Notably, we achieve an average relative performance improvement of 11.8% in reference-free metrics over the previous SOTA. Doctors prefer uMedSum's summaries 6 times more than previous SOTA in difficult cases where there are chances of confabulations or missing information. These results highlight uMedSum's effectiveness and generalizability across various datasets and metrics, marking a significant advancement in medical summarization.

Yiming Chen, Xianghu Yue, Xiaoxue Gao et al.

Various audio-LLMs (ALLMs) have been explored recently for tackling different audio tasks simultaneously using a single, unified model. While existing evaluations of ALLMs primarily focus on single-audio tasks, real-world applications often involve processing multiple audio streams simultaneously. To bridge this gap, we propose the first multi-audio evaluation (MAE) benchmark that consists of 20 datasets from 11 multi-audio tasks encompassing both speech and sound scenarios. Comprehensive experiments on MAE demonstrate that the existing ALLMs, while being powerful in comprehending primary audio elements in individual audio inputs, struggling to handle multi-audio scenarios. To this end, we propose a novel multi-audio-LLM (MALLM) to capture audio context among multiple similar audios using discriminative learning on our proposed synthetic data. The results demonstrate that the proposed MALLM outperforms all baselines and achieves high data efficiency using synthetic data without requiring human annotations. The proposed MALLM opens the door for ALLMs towards multi-audio processing era and brings us closer to replicating human auditory capabilities in machines.

Yu Cheng, Bo Wang, Robby T. Tan

In 3D human shape and pose estimation from a monocular video, models trained with limited labeled data cannot generalize well to videos with occlusion, which is common in the wild videos. The recent human neural rendering approaches focusing on novel view synthesis initialized by the off-the-shelf human shape and pose methods have the potential to correct the initial human shape. However, the existing methods have some drawbacks such as, erroneous in handling occlusion, sensitive to inaccurate human segmentation, and ineffective loss computation due to the non-regularized opacity field. To address these problems, we introduce ORTexME, an occlusion-robust temporal method that utilizes temporal information from the input video to better regularize the occluded body parts. While our ORTexME is based on NeRF, to determine the reliable regions for the NeRF ray sampling, we utilize our novel average texture learning approach to learn the average appearance of a person, and to infer a mask based on the average texture. In addition, to guide the opacity-field updates in NeRF to suppress blur and noise, we propose the use of human body mesh. The quantitative evaluation demonstrates that our method achieves significant improvement on the challenging multi-person 3DPW dataset, where our method achieves 1.8 P-MPJPE error reduction. The SOTA rendering-based methods fail and enlarge the error up to 5.6 on the same dataset.

Shuwei Li, Lei Tan, Robby T. Tan

Day-to-night unpaired image translation is important to downstream tasks but remains challenging due to large appearance shifts and the lack of direct pixel-level supervision. Existing methods often introduce semantic hallucinations, where objects from target classes such as traffic signs and vehicles, as well as man-made light effects, are incorrectly synthesized. These hallucinations significantly degrade downstream performance. We propose a novel framework that detects and suppresses hallucinations of target-class features during unpaired translation. To detect hallucination, we design a dual-head discriminator that additionally performs semantic segmentation to identify hallucinated content in background regions. To suppress these hallucinations, we introduce class-specific prototypes, constructed by aggregating features of annotated target-domain objects, which act as semantic anchors for each class. Built upon a Schrodinger Bridge-based translation model, our framework performs iterative refinement, where detected hallucination features are explicitly pushed away from class prototypes in feature space, thus preserving object semantics across the translation trajectory.Experiments show that our method outperforms existing approaches both qualitatively and quantitatively. On the BDD100K dataset, it improves mAP by 15.5% for day-to-night domain adaptation, with a notable 31.7% gain for classes such as traffic lights that are prone to hallucinations.

Lei Tan, Shuwei Li, Mohan Kankanhalli et al.

The rapid emergence of image synthesis models poses challenges to the generalization of AI-generated image detectors. However, existing methods often rely on model-specific features, leading to overfitting and poor generalization. In this paper, we introduce the Multi-Cue Aggregation Network (MCAN), a novel framework that integrates different yet complementary cues in a unified network. MCAN employs a mixture-of-encoders adapter to dynamically process these cues, enabling more adaptive and robust feature representation. Our cues include the input image itself, which represents the overall content, and high-frequency components that emphasize edge details. Additionally, we introduce a Chromatic Inconsistency (CI) cue, which normalizes intensity values and captures noise information introduced during the image acquisition process in real images, making these noise patterns more distinguishable from those in AI-generated content. Unlike prior methods, MCAN's novelty lies in its unified multi-cue aggregation framework, which integrates spatial, frequency-domain, and chromaticity-based information for enhanced representation learning. These cues are intrinsically more indicative of real images, enhancing cross-model generalization. Extensive experiments on the GenImage, Chameleon, and UniversalFakeDetect benchmark validate the state-of-the-art performance of MCAN. In the GenImage dataset, MCAN outperforms the best state-of-the-art method by up to 7.4% in average ACC across eight different image generators.

Yushi Cao, Yiming Chen, Hongchao Jiang et al.

Predicting future clinical outcomes from electronic health records (EHR) remains challenging due to the complexity and heterogeneity of patient data. LLMs have shown strong potential for such predictive tasks, yet existing approaches mainly focus on enhancing medical knowledge through distillation or RAG while relying on the model's internal ability to interpret contextual information. In this work, we present ReMedi (Reasoner for Medical Clinical Prediction), a framework for improving clinical outcome prediction from EHR. ReMedi generates rationale-answer pairs using a challenging sample regeneration mechanism for complex clinical questions, which leverages ground-truth answers as hints to enhance reasoning for further fine-tuning and preference tuning. ReMedi integrates ground-truth outcome guidance into the preference data construction loop, regenerating rationale-answer variants. By tuning on these rationale-answer pairs, the model improves its predictive performance. Experiments on multiple EHR prediction tasks demonstrate substantial gains of up to 19.9 percent over state-of-the-art baselines in terms of F1 score, underscoring ReMedi's effectiveness in real-world clinical prediction.

Beibei Lin, Xiao Cao, Jingyuan Guo et al.

Existing 3DGS methods effectively render high-quality novel views in clear-day scenes. However, they struggle with night scenes, particularly in glow regions, due to the lack of structural features such as textures and edges, which are key cues for splatting-based reconstruction. To address this problem, we leverage a diffusion model and a Vision Foundation Model (VFM) to compensate for missing structural cues. Our method consists of two key novel ideas: semantic feature generation and novel-view semantic learning. First, semantic feature generation produces high-quality semantic features as implicit structural cues for novel views. Specifically, a diffusion model synthesizes novel views with unknown camera poses from training views, while a VFM evaluates their quality. Once high-quality novel views are identified, the VFM extracts robust features to construct the semantic feature bank. Second, novel-view semantic learning enables 3DGS to optimize rendered novel views without requiring ground truth. It achieves this by extracting semantic features from a rendered novel view, searching the feature bank for the most similar features, and minimizing their distance. This process enforces implicit structural constraints, ensuring semantically coherent, artifact-free rendered views. Extensive experiments demonstrate the effectiveness of our GlowGS in generating semantically accurate 3D views, showing significant improvements over existing methods.

Lemeng Wang, Qinqian Lei, Vidhi Bakshi et al.

Human-object interaction (HOI) detection aims to detect interactions between humans and objects in images. While recent advances have improved performance on existing benchmarks, their evaluations mainly focus on overall prediction accuracy and provide limited insight into the underlying causes of model failures. In particular, modern models often struggle in complex scenes involving multiple people and rare interaction combinations. In this work, we present a study to better understand the failure modes of two-stage HOI models, which form the basis of many current HOI detection approaches. Rather than constructing a large-scale benchmark, we instead decompose HOI detection into multiple interpretable perspectives and analyze model behavior across these dimensions to study different types of failure patterns. We curate a subset of images from an existing HOI dataset organized by human-object-interaction configurations (e.g., multi-person interactions and object sharing), and analyze model behavior under these configurations to examine different failure modes. This design allows us to analyze how these HOI models behave under different scene compositions and why their predictions fail. Importantly, high overall benchmark performance does not necessarily reflect robust visual reasoning about human-object relationships. We hope that this study can provide useful insights into the limitations of HOI models and offer observations for future research in this area.

Qinqian Lei, Bo Wang, Robby T. Tan

Detecting Human-Object Interactions (HOI) in zero-shot settings, where models must handle unseen classes, poses significant challenges. Existing methods that rely on aligning visual encoders with large Vision-Language Models (VLMs) to tap into the extensive knowledge of VLMs, require large, computationally expensive models and encounter training difficulties. Adapting VLMs with prompt learning offers an alternative to direct alignment. However, fine-tuning on task-specific datasets often leads to overfitting to seen classes and suboptimal performance on unseen classes, due to the absence of unseen class labels. To address these challenges, we introduce a novel prompt learning-based framework for Efficient Zero-Shot HOI detection (EZ-HOI). First, we introduce Large Language Model (LLM) and VLM guidance for learnable prompts, integrating detailed HOI descriptions and visual semantics to adapt VLMs to HOI tasks. However, because training datasets contain seen-class labels alone, fine-tuning VLMs on such datasets tends to optimize learnable prompts for seen classes instead of unseen ones. Therefore, we design prompt learning for unseen classes using information from related seen classes, with LLMs utilized to highlight the differences between unseen and related seen classes. Quantitative evaluations on benchmark datasets demonstrate that our EZ-HOI achieves state-of-the-art performance across various zero-shot settings with only 10.35% to 33.95% of the trainable parameters compared to existing methods. Code is available at https://github.com/ChelsieLei/EZ-HOI.

Xiao Cao, Beibei Lin, Bo Wang et al.

Sparse view NeRF is challenging because limited input images lead to an under constrained optimization problem for volume rendering. Existing methods address this issue by relying on supplementary information, such as depth maps. However, generating this supplementary information accurately remains problematic and often leads to NeRF producing images with undesired artifacts. To address these artifacts and enhance robustness, we propose SSNeRF, a sparse view semi supervised NeRF method based on a teacher student framework. Our key idea is to challenge the NeRF module with progressively severe sparse view degradation while providing high confidence pseudo labels. This approach helps the NeRF model become aware of noise and incomplete information associated with sparse views, thus improving its robustness. The novelty of SSNeRF lies in its sparse view specific augmentations and semi supervised learning mechanism. In this approach, the teacher NeRF generates novel views along with confidence scores, while the student NeRF, perturbed by the augmented input, learns from the high confidence pseudo labels. Our sparse view degradation augmentation progressively injects noise into volume rendering weights, perturbs feature maps in vulnerable layers, and simulates sparse view blurriness. These augmentation strategies force the student NeRF to recognize degradation and produce clearer rendered views. By transferring the student's parameters to the teacher, the teacher gains increased robustness in subsequent training iterations. Extensive experiments demonstrate the effectiveness of our SSNeRF in generating novel views with less sparse view degradation. We will release code upon acceptance.

Xin Zhang, Robby T. Tan

Vision Foundation Models (VFMs) and Vision-Language Models (VLMs) have gained traction in Domain Generalized Semantic Segmentation (DGSS) due to their strong generalization capabilities. However, existing DGSS methods often rely exclusively on either VFMs or VLMs, overlooking their complementary strengths. VFMs (e.g., DINOv2) excel at capturing fine-grained features, while VLMs (e.g., CLIP) provide robust text alignment but struggle with coarse granularity. Despite their complementary strengths, effectively integrating VFMs and VLMs with attention mechanisms is challenging, as the increased patch tokens complicate long-sequence modeling. To address this, we propose MFuser, a novel Mamba-based fusion framework that efficiently combines the strengths of VFMs and VLMs while maintaining linear scalability in sequence length. MFuser consists of two key components: MVFuser, which acts as a co-adapter to jointly fine-tune the two models by capturing both sequential and spatial dynamics; and MTEnhancer, a hybrid attention-Mamba module that refines text embeddings by incorporating image priors. Our approach achieves precise feature locality and strong text alignment without incurring significant computational overhead. Extensive experiments demonstrate that MFuser significantly outperforms state-of-the-art DGSS methods, achieving 68.20 mIoU on synthetic-to-real and 71.87 mIoU on real-to-real benchmarks. The code is available at https://github.com/devinxzhang/MFuser.

Maja Noack, Qinqian Lei, Taipeng Tian et al.

Open-vocabulary human-object interaction (HOI) detection is a step towards building scalable systems that generalize to unseen interactions in real-world scenarios and support grounded multimodal systems that reason about human-object relationships. However, standard evaluation metrics, such as mean Average Precision (mAP), treat HOI classes as discrete categorical labels and fail to credit semantically valid but lexically different predictions (e.g., "lean on couch" vs. "sit on couch"), limiting their applicability for evaluating open-vocabulary predictions that go beyond any predefined set of HOI labels. We introduce SHOE (Semantic HOI Open-Vocabulary Evaluation), a new evaluation framework that incorporates semantic similarity between predicted and ground-truth HOI labels. SHOE decomposes each HOI prediction into its verb and object components, estimates their semantic similarity using the average of multiple large language models (LLMs), and combines them into a similarity score to evaluate alignment beyond exact string match. This enables a flexible and scalable evaluation of both existing HOI detection methods and open-ended generative models using standard benchmarks such as HICO-DET. Experimental results show that SHOE scores align more closely with human judgments than existing metrics, including LLM-based and embedding-based baselines, achieving an agreement of 85.73% with the average human ratings. Our work underscores the need for semantically grounded HOI evaluation that better mirrors human understanding of interactions. We will release our evaluation metric to the public to facilitate future research.

Qinqian Lei, Bo Wang, Robby T. Tan

Zero-shot human-object interaction (HOI) detection remains a challenging task, particularly in generalizing to unseen actions. Existing methods address this challenge by tapping Vision-Language Models (VLMs) to access knowledge beyond the training data. However, they either struggle to distinguish actions involving the same object or demonstrate limited generalization to unseen classes. In this paper, we introduce HOLa (Zero-Shot HOI Detection with Low-Rank Decomposed VLM Feature Adaptation), a novel approach that both enhances generalization to unseen classes and improves action distinction. In training, HOLa decomposes VLM text features for given HOI classes via low-rank factorization, producing class-shared basis features and adaptable weights. These features and weights form a compact HOI representation that preserves shared information across classes, enhancing generalization to unseen classes. Subsequently, we refine action distinction by adapting weights for each HOI class and introducing human-object tokens to enrich visual interaction representations. To further distinguish unseen actions, we guide the weight adaptation with LLM-derived action regularization. Experimental results show that our method sets a new state-of-the-art across zero-shot HOI settings on HICO-DET, achieving an unseen-class mAP of 27.91 in the unseen-verb setting. Our code is available at https://github.com/ChelsieLei/HOLa.

Jialun Liu, Yukuo Ma, Xiao Cao et al.

Recent advances in diffusion-based video generation have substantially improved visual fidelity and temporal coherence. However, most existing approaches remain task-specific and rely primarily on textual instructions, limiting their ability to handle multimodal inputs, contextual references, and diverse video generation and editing scenarios within a unified framework. Moreover, many video editing methods depend on carefully engineered pipelines tailored to individual operations, which hinders scalability and composability. In this paper, we propose Tele-Omni, a unified multimodal framework for video generation and editing that follows multimodal instructions, including text, images, and reference videos, within a single model. Tele-Omni leverages pretrained multimodal large language models to parse heterogeneous instructions and infer structured generation or editing intents, while diffusion-based generators perform high-quality video synthesis conditioned on these structured signals. To enable joint training across heterogeneous video tasks, we introduce a task-aware data processing pipeline that unifies multimodal inputs into a structured instruction format while preserving task-specific constraints. Tele-Omni supports a wide range of video-centric tasks, including text-to-video generation, image-to-video generation, first-last-frame video generation, in-context video generation, and in-context video editing. By decoupling instruction parsing from video synthesis and combining it with task-aware data design, Tele-Omni achieves flexible multimodal control while maintaining strong temporal coherence and visual consistency. Experimental results demonstrate that Tele-Omni achieves competitive performance across multiple tasks.

Heyuan Li, Bo Wang, Yu Cheng et al.

Sensitivity to severe occlusion and large view angles limits the usage scenarios of the existing monocular 3D dense face alignment methods. The state-of-the-art 3DMM-based method, directly regresses the model's coefficients, underutilizing the low-level 2D spatial and semantic information, which can actually offer cues for face shape and orientation. In this work, we demonstrate how modeling 3D facial geometry in image and model space jointly can solve the occlusion and view angle problems. Instead of predicting the whole face directly, we regress image space features in the visible facial region by dense prediction first. Subsequently, we predict our model's coefficients based on the regressed feature of the visible regions, leveraging the prior knowledge of whole face geometry from the morphable models to complete the invisible regions. We further propose a fusion network that combines the advantages of both the image and model space predictions to achieve high robustness and accuracy in unconstrained scenarios. Thanks to the proposed fusion module, our method is robust not only to occlusion and large pitch and roll view angles, which is the benefit of our image space approach, but also to noise and large yaw angles, which is the benefit of our model space method. Comprehensive evaluations demonstrate the superior performance of our method compared with the state-of-the-art methods. On the 3D dense face alignment task, we achieve 3.80% NME on the AFLW2000-3D dataset, which outperforms the state-of-the-art method by 5.5%. Code is available at https://github.com/lhyfst/DSFNet.

Yiming Chen, Xianghu Yue, Chen Zhang et al.

Building on the success of large language models (LLMs), recent advancements such as GPT-4o have enabled real-time speech interactions through LLM-based voice assistants, offering a significantly improved user experience compared to traditional text-based interactions. However, the absence of benchmarks designed to evaluate these speech interaction capabilities has hindered progress of LLM-based voice assistants development. Current evaluations focus primarily on automatic speech recognition (ASR) or general knowledge evaluation with clean speeches, neglecting the more intricate, real-world scenarios that involve diverse speaker characteristics, environmental and content factors. To address this, we introduce VoiceBench, the first benchmark designed to provide a multi-faceted evaluation of LLM-based voice assistants. VoiceBench also includes both real and synthetic spoken instructions that incorporate the above three key real-world variations. Extensive experiments reveal the limitations of current LLM-based voice assistant models and offer valuable insights for future research and development in this field.

Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli et al.

Domain adaptive object detection aims to adapt detection models to domains where annotated data is unavailable. Existing methods have been proposed to address the domain gap using the semi-supervised student-teacher framework. However, a fundamental issue arises from the class imbalance in the labelled training set, which can result in inaccurate pseudo-labels. The relationship between classes, especially where one class is a majority and the other minority, has a large impact on class bias. We propose Class-Aware Teacher (CAT) to address the class bias issue in the domain adaptation setting. In our work, we approximate the class relationships with our Inter-Class Relation module (ICRm) and exploit it to reduce the bias within the model. In this way, we are able to apply augmentations to highly related classes, both inter- and intra-domain, to boost the performance of minority classes while having minimal impact on majority classes. We further reduce the bias by implementing a class-relation weight to our classification loss. Experiments conducted on various datasets and ablation studies show that our method is able to address the class bias in the domain adaptation setting. On the Cityscapes to Foggy Cityscapes dataset, we attained a 52.5 mAP, a substantial improvement over the 51.2 mAP achieved by the state-of-the-art method.

Xin Li, Yeying Jin, Xin Jin et al.

This paper reviews the NTIRE 2025 Challenge on Day and Night Raindrop Removal for Dual-Focused Images. This challenge received a wide range of impressive solutions, which are developed and evaluated using our collected real-world Raindrop Clarity dataset. Unlike existing deraining datasets, our Raindrop Clarity dataset is more diverse and challenging in degradation types and contents, which includes day raindrop-focused, day background-focused, night raindrop-focused, and night background-focused degradations. This dataset is divided into three subsets for competition: 14,139 images for training, 240 images for validation, and 731 images for testing. The primary objective of this challenge is to establish a new and powerful benchmark for the task of removing raindrops under varying lighting and focus conditions. There are a total of 361 participants in the competition, and 32 teams submitting valid solutions and fact sheets for the final testing phase. These submissions achieved state-of-the-art (SOTA) performance on the Raindrop Clarity dataset. The project can be found at https://lixinustc.github.io/CVPR-NTIRE2025-RainDrop-Competition.github.io/.

Beibei Lin, Yeying Jin, Wending Yan et al.

Masked autoencoder (MAE) shows that severe augmentation during training produces robust representations for high-level tasks. This paper brings the MAE-like framework to nighttime image enhancement, demonstrating that severe augmentation during training produces strong network priors that are resilient to real-world night haze degradations. We propose a novel nighttime image dehazing method with self-prior learning. Our main novelty lies in the design of severe augmentation, which allows our model to learn robust priors. Unlike MAE that uses masking, we leverage two key challenging factors of nighttime images as augmentation: light effects and noise. During training, we intentionally degrade clear images by blending them with light effects as well as by adding noise, and subsequently restore the clear images. This enables our model to learn clear background priors. By increasing the noise values to approach as high as the pixel intensity values of the glow and light effect blended images, our augmentation becomes severe, resulting in stronger priors. While our self-prior learning is considerably effective in suppressing glow and revealing details of background scenes, in some cases, there are still some undesired artifacts that remain, particularly in the forms of over-suppression. To address these artifacts, we propose a self-refinement module based on the semi-supervised teacher-student framework. Our NightHaze, especially our MAE-like self-prior learning, shows that models trained with severe augmentation effectively improve the visibility of input haze images, approaching the clarity of clear nighttime images. Extensive experiments demonstrate that our NightHaze achieves state-of-the-art performance, outperforming existing nighttime image dehazing methods by a substantial margin of 15.5% for MUSIQ and 23.5% for ClipIQA.

Weilong Yan, Ming Li, Haipeng Li et al.

Self-supervised depth estimation from monocular cameras in diverse outdoor conditions, such as daytime, rain, and nighttime, is challenging due to the difficulty of learning universal representations and the severe lack of labeled real-world adverse data. Previous methods either rely on synthetic inputs and pseudo-depth labels or directly apply daytime strategies to adverse conditions, resulting in suboptimal results. In this paper, we present the first synthetic-to-real robust depth estimation framework, incorporating motion and structure priors to capture real-world knowledge effectively. In the synthetic adaptation, we transfer motion-structure knowledge inside cost volumes for better robust representation, using a frozen daytime model to train a depth estimator in synthetic adverse conditions. In the innovative real adaptation, which targets to fix synthetic-real gaps, models trained earlier identify the weather-insensitive regions with a designed consistency-reweighting strategy to emphasize valid pseudo-labels. We introduce a new regularization by gathering explicit depth distributions to constrain the model when facing real-world data. Experiments show that our method outperforms the state-of-the-art across diverse conditions in multi-frame and single-frame evaluations. We achieve improvements of 7.5% and 4.3% in AbsRel and RMSE on average for nuScenes and Robotcar datasets (daytime, nighttime, rain). In zero-shot evaluation of DrivingStereo (rain, fog), our method generalizes better than the previous ones.

Xin Yang, Wending Yan, Yuan Yuan et al.

Semantic segmentation's performance is often compromised when applied to unlabeled adverse weather conditions. Unsupervised domain adaptation is a potential approach to enhancing the model's adaptability and robustness to adverse weather. However, existing methods encounter difficulties when sequentially adapting the model to multiple unlabeled adverse weather conditions. They struggle to acquire new knowledge while also retaining previously learned knowledge.To address these problems, we propose a semantic segmentation method for multiple adverse weather conditions that incorporates adaptive knowledge acquisition, pseudolabel blending, and weather composition replay. Our adaptive knowledge acquisition enables the model to avoid learning from extreme images that could potentially cause the model to forget. In our approach of blending pseudo-labels, we not only utilize the current model but also integrate the previously learned model into the ongoing learning process. This collaboration between the current teacher and the previous model enhances the robustness of the pseudo-labels for the current target. Our weather composition replay mechanism allows the model to continuously refine its previously learned weather information while simultaneously learning from the new target domain. Our method consistently outperforms the stateof-the-art methods, and obtains the best performance with averaged mIoU (%) of 65.7 and the lowest forgetting (%) of 3.6 against 60.1 and 11.3, on the ACDC datasets for a four-target continual multi-target domain adaptation.

Xin Zhang, Jinheng Xie, Yuan Yuan et al.

Unsupervised object discovery and localization aims to detect or segment objects in an image without any supervision. Recent efforts have demonstrated a notable potential to identify salient foreground objects by utilizing self-supervised transformer features. However, their scopes only build upon patch-level features within an image, neglecting region/image-level and cross-image relationships at a broader scale. Moreover, these methods cannot differentiate various semantics from multiple instances. To address these problems, we introduce Hierarchical mErging framework via contrAstive grouPing (HEAP). Specifically, a novel lightweight head with cross-attention mechanism is designed to adaptively group intra-image patches into semantically coherent regions based on correlation among self-supervised features. Further, to ensure the distinguishability among various regions, we introduce a region-level contrastive clustering loss to pull closer similar regions across images. Also, an image-level contrastive loss is present to push foreground and background representations apart, with which foreground objects and background are accordingly discovered. HEAP facilitates efficient hierarchical image decomposition, which contributes to more accurate object discovery while also enabling differentiation among objects of various classes. Extensive experimental results on semantic segmentation retrieval, unsupervised object discovery, and saliency detection tasks demonstrate that HEAP achieves state-of-the-art performance.

Hongchao Jiang, Yiming Chen, Yushi Cao et al.

Large Language Models (LLMs) have significantly advanced the state-of-the-art in various coding tasks. Beyond directly answering user queries, LLMs can also serve as judges, assessing and comparing the quality of responses generated by other models. Such an evaluation capability is crucial both for benchmarking different LLMs and for improving response quality through response ranking. However, despite the growing adoption of the LLM-as-a-Judge paradigm, its effectiveness in coding scenarios remains underexplored due to the absence of dedicated benchmarks. To address this gap, we introduce CodeJudgeBench, a benchmark explicitly designed to evaluate the performance of LLM-as-a-Judge models across three critical coding tasks: code generation, code repair, and unit test generation. Through comprehensive benchmarking of 26 LLM-as-a-Judge models, we find that recent thinking models significantly outperform non-thinking models on our carefully designed code judging tasks. Notably, even relatively small thinking models, such as Qwen3-8B, can outperform specially trained LLM-as-a-Judge models up to 70B in size. Nevertheless, all models still exhibit significant randomness in their judgment of coding tasks. For pairwise judging tasks, simply changing the order in which responses are presented can substantially impact accuracy. In addition, when judging code and unit tests written by different LLMs, LLM-as-a-Judge models also show variance in performance. This sensitivity raises concerns about the reliability and consistency of LLM-as-a-Judge in coding scenarios. Lastly, we study optimal prompting strategies for LLM-as-a-Judge. We find that using pair-wise comparison outperforms scalar point-wise judging. Furthermore, retaining comments and reasoning in the full, unprocessed LLM response leads to improved judge performance.

Qinqian Lei, Bo Wang, Robby T. Tan

Detecting human-object interactions (HOI) in a few-shot setting remains a challenge. Existing meta-learning methods struggle to extract representative features for classification due to the limited data, while existing few-shot HOI models rely on HOI text labels for classification. Moreover, some query images may display visual similarity to those outside their class, such as similar backgrounds between different HOI classes. This makes learning more challenging, especially with limited samples. Bongard-HOI (Jiang et al. 2022) epitomizes this HOI few-shot problem, making it the benchmark we focus on in this paper. In our proposed method, we introduce novel label-uncertain query augmentation techniques to enhance the diversity of the query inputs, aiming to distinguish the positive HOI class from the negative ones. As these augmented inputs may or may not have the same class label as the original inputs, their class label is unknown. Those belonging to a different class become hard samples due to their visual similarity to the original ones. Additionally, we introduce a novel pseudo-label generation technique that enables a mean teacher model to learn from the augmented label-uncertain inputs. We propose to augment the negative support set for the student model to enrich the semantic information, fostering diversity that challenges and enhances the student's learning. Experimental results demonstrate that our method sets a new state-of-the-art (SOTA) performance by achieving 68.74% accuracy on the Bongard-HOI benchmark, a significant improvement over the existing SOTA of 66.59%. In our evaluation on HICO-FS, a more general few-shot recognition dataset, our method achieves 73.27% accuracy, outperforming the previous SOTA of 71.20% in the 5-way 5-shot task.

Beibei Lin, Tingting Chen, Robby T. Tan

Reference-driven image completion, which restores missing regions in a target view using additional images, is particularly challenging when the target view differs significantly from the references. Existing generative methods rely solely on diffusion priors and, without geometric cues such as camera pose or depth, often produce misaligned or implausible content. We propose GeoComplete, a novel framework that incorporates explicit 3D structural guidance to enforce geometric consistency in the completed regions, setting it apart from prior image-only approaches. GeoComplete introduces two key ideas: conditioning the diffusion process on projected point clouds to infuse geometric information, and applying target-aware masking to guide the model toward relevant reference cues. The framework features a dual-branch diffusion architecture. One branch synthesizes the missing regions from the masked target, while the other extracts geometric features from the projected point cloud. Joint self-attention across branches ensures coherent and accurate completion. To address regions visible in references but absent in the target, we project the target view into each reference to detect occluded areas, which are then masked during training. This target-aware masking directs the model to focus on useful cues, enhancing performance in difficult scenarios. By integrating a geometry-aware dual-branch diffusion architecture with a target-aware masking strategy, GeoComplete offers a unified and robust solution for geometry-conditioned image completion. Experiments show that GeoComplete achieves a 17.1 PSNR improvement over state-of-the-art methods, significantly boosting geometric accuracy while maintaining high visual quality.

Yiming Chen, Chen Zhang, Danqing Luo et al.

The automatic evaluation of natural language generation (NLG) systems presents a long-lasting challenge. Recent studies have highlighted various neural metrics that align well with human evaluations. Yet, the robustness of these evaluators against adversarial perturbations remains largely under-explored due to the unique challenges in obtaining adversarial data for different NLG evaluation tasks. To address the problem, we introduce AdvEval, a novel black-box adversarial framework against NLG evaluators. AdvEval is specially tailored to generate data that yield strong disagreements between human and victim evaluators. Specifically, inspired by the recent success of large language models (LLMs) in text generation and evaluation, we adopt strong LLMs as both the data generator and gold evaluator. Adversarial data are automatically optimized with feedback from the gold and victim evaluator. We conduct experiments on 12 victim evaluators and 11 NLG datasets, spanning tasks including dialogue, summarization, and question evaluation. The results show that AdvEval can lead to significant performance degradation of various victim metrics, thereby validating its efficacy.

Xiao Cao, Beibei Lin, Bo Wang et al.

3D texture swapping allows for the customization of 3D object textures, enabling efficient and versatile visual transformations in 3D editing. While no dedicated method exists, adapted 2D editing and text-driven 3D editing approaches can serve this purpose. However, 2D editing requires frame-by-frame manipulation, causing inconsistencies across views, while text-driven 3D editing struggles to preserve texture characteristics from reference images. To tackle these challenges, we introduce 3DSwapping, a 3D texture swapping method that integrates: 1) progressive generation, 2) view-consistency gradient guidance, and 3) prompt-tuned gradient guidance. To ensure view consistency, our progressive generation process starts by editing a single reference image and gradually propagates the edits to adjacent views. Our view-consistency gradient guidance further reinforces consistency by conditioning the generation model on feature differences between consistent and inconsistent outputs. To preserve texture characteristics, we introduce prompt-tuning-based gradient guidance, which learns a token that precisely captures the difference between the reference image and the 3D object. This token then guides the editing process, ensuring more consistent texture preservation across views. Overall, 3DSwapping integrates these novel strategies to achieve higher-fidelity texture transfer while preserving structural coherence across multiple viewpoints. Extensive qualitative and quantitative evaluations confirm that our three novel components enable convincing and effective 2D texture swapping for 3D objects. Code will be available upon acceptance.

Vijay Prakash Dwivedi, Viktor Schlegel, Andy T. Liu et al.

Large Language Models (LLMs) have demonstrated remarkable performance across various domains, including healthcare. However, their ability to effectively represent structured non-textual data, such as the alphanumeric medical codes used in records like ICD-10 or SNOMED-CT, is limited and has been particularly exposed in recent research. This paper examines the challenges LLMs face in processing medical codes due to the shortcomings of current tokenization methods. As a result, we introduce the UniStruct architecture to design a multimodal medical foundation model of unstructured text and structured data, which addresses these challenges by adapting subword tokenization techniques specifically for the structured medical codes. Our approach is validated through model pre-training on both an extensive internal medical database and a public repository of structured medical records. Trained on over 1 billion tokens on the internal medical database, the proposed model achieves up to a 23% improvement in evaluation metrics, with around 2% gain attributed to our proposed tokenization. Additionally, when evaluated on the EHRSHOT public benchmark with a 1/1000 fraction of the pre-training data, the UniStruct model improves performance on over 42% of the downstream tasks. Our approach not only enhances the representation and generalization capabilities of patient-centric models but also bridges a critical gap in representation learning models' ability to handle complex structured medical data, alongside unstructured text.

Weilong Yan, Xin Zhang, Robby T. Tan

Monocular depth estimation under adverse weather conditions (e.g.\ rain, fog, snow, and nighttime) remains highly challenging due to the lack of reliable ground truth and the difficulty of learning from unlabeled real-world data. Existing methods often rely on synthetic adverse data with pseudo-labels, which suffer from domain gaps, or employ self-supervised learning, which violates photometric assumptions in adverse scenarios. In this work, we propose to achieve weather-generalized depth estimation by Parameter-Efficient Fine-Tuning (PEFT) of Vision Foundation Models (VFMs), using only a small amount of high-visibility (normal) data. While PEFT has shown strong performance in semantic tasks such as segmentation, it remains underexplored for geometry -- centric tasks like depth estimation -- especially in terms of balancing effective adaptation with the preservation of pretrained knowledge. To this end, we introduce the Selecting-Tuning-Maintaining (STM) strategy, which structurally decomposes the pretrained weights of VFMs based on two kinds of effective ranks (entropy-rank and stable-rank). In the tuning phase, we adaptively select the proper rank number as well as the task-aware singular directions for initialization, based on the entropy-rank and full-tuned weight; while in the maintaining stage, we enforce a principal direction regularization based on the stable-rank. This design guarantees flexible task adaptation while preserving the strong generalization capability of the pretrained VFM. Extensive experiments on four real-world benchmarks across diverse weather conditions demonstrate that STM not only outperforms existing PEFT methods and full fine-tuning but also surpasses methods trained with adverse synthetic data, and even the depth foundation model

Qinqian Lei, Bo Wang, Robby T. Tan

Human-object interaction (HOI) detection has traditionally been approached with task-specific models, sometimes augmented by early vision-language models (VLMs) such as CLIP. With the rise of large, generative VLMs, however, a natural question emerges: can standalone VLMs effectively perform HOI detection, and how do they compare to specialized HOI methods? Addressing this requires a benchmarking dataset and protocol that support both paradigms. Existing benchmarks such as HICO-DET were developed before modern VLMs and rely on exact label matching. This clashes with generative outputs, which may yield multiple equally valid interpretations. For example, in a single image, a person mid-motion with a frisbee might plausibly be described as 'throwing' or 'catching', yet only one is annotated as correct. Such rigid evaluation penalizes valid predictions from both VLMs and HOI-specific methods, but disproportionately underestimates VLM performance because their outputs are less constrained. We introduce a new benchmarking dataset that reformulates HOI detection as a multiple-answer multiple-choice task. It emphasizes challenging scenarios by (i) including a higher proportion of multi-person scenes where individuals perform different interactions, (ii) removing overly simple cases, and (iii) curating hard negative choices. This makes the benchmark more challenging than prior HOI datasets, while still supporting systematic evaluation of both standalone VLMs and HOI-specific methods under a unified protocol. Our results show that large VLMs already surpass state-of-the-art HOI-specific methods across most metrics, while analysis further uncovers key limitations: VLMs often misattribute surrounding people's interactions to the target person and struggle in complex multi-person or occluded scenarios.

Mikhail Kennerley, Angelica Aviles-Rivero, Carola-Bibiane Schönlieb et al.

Combining multiple object detection datasets offers a path to improved generalisation but is hindered by inconsistencies in class semantics and bounding box annotations. Some methods to address this assume shared label taxonomies and address only spatial inconsistencies; others require manual relabelling, or produce a unified label space, which may be unsuitable when a fixed target label space is required. We propose Label-Aligned Transfer (LAT), a label transfer framework that systematically projects annotations from diverse source datasets into the label space of a target dataset. LAT begins by training dataset-specific detectors to generate pseudo-labels, which are then combined with ground-truth annotations via a Privileged Proposal Generator (PPG) that replaces the region proposal network in two-stage detectors. To further refine region features, a Semantic Feature Fusion (SFF) module injects class-aware context and features from overlapping proposals using a confidence-weighted attention mechanism. This pipeline preserves dataset-specific annotation granularity while enabling many-to-one label space transfer across heterogeneous datasets, resulting in a semantically and spatially aligned representation suitable for training a downstream detector. LAT thus jointly addresses both class-level misalignments and bounding box inconsistencies without relying on shared label spaces or manual annotations. Across multiple benchmarks, LAT demonstrates consistent improvements in target-domain detection performance, achieving gains of up to +4.8AP over semi-supervised baselines.