Haoqian Wang

Yuanhao Cai, Hao Bian, Jing Lin et al. · eth-zurich

When enhancing low-light images, many deep learning algorithms are based on the Retinex theory. However, the Retinex model does not consider the corruptions hidden in the dark or introduced by the light-up process. Besides, these methods usually require a tedious multi-stage training pipeline and rely on convolutional neural networks, showing limitations in capturing long-range dependencies. In this paper, we formulate a simple yet principled One-stage Retinex-based Framework (ORF). ORF first estimates the illumination information to light up the low-light image and then restores the corruption to produce the enhanced image. We design an Illumination-Guided Transformer (IGT) that utilizes illumination representations to direct the modeling of non-local interactions of regions with different lighting conditions. By plugging IGT into ORF, we obtain our algorithm, Retinexformer. Comprehensive quantitative and qualitative experiments demonstrate that our Retinexformer significantly outperforms state-of-the-art methods on thirteen benchmarks. The user study and application on low-light object detection also reveal the latent practical values of our method. Code, models, and results are available at https://github.com/caiyuanhao1998/Retinexformer

Yuanhao Cai, Jing Lin, Zudi Lin et al. · harvard

Existing leading methods for spectral reconstruction (SR) focus on designing deeper or wider convolutional neural networks (CNNs) to learn the end-to-end mapping from the RGB image to its hyperspectral image (HSI). These CNN-based methods achieve impressive restoration performance while showing limitations in capturing the long-range dependencies and self-similarity prior. To cope with this problem, we propose a novel Transformer-based method, Multi-stage Spectral-wise Transformer (MST++), for efficient spectral reconstruction. In particular, we employ Spectral-wise Multi-head Self-attention (S-MSA) that is based on the HSI spatially sparse while spectrally self-similar nature to compose the basic unit, Spectral-wise Attention Block (SAB). Then SABs build up Single-stage Spectral-wise Transformer (SST) that exploits a U-shaped structure to extract multi-resolution contextual information. Finally, our MST++, cascaded by several SSTs, progressively improves the reconstruction quality from coarse to fine. Comprehensive experiments show that our MST++ significantly outperforms other state-of-the-art methods. In the NTIRE 2022 Spectral Reconstruction Challenge, our approach won the First place. Code and pre-trained models are publicly available at https://github.com/caiyuanhao1998/MST-plus-plus.

Xiaowan Hu, Yuanhao Cai, Jing Lin et al. · tsinghua

The rapid development of deep learning provides a better solution for the end-to-end reconstruction of hyperspectral image (HSI). However, existing learning-based methods have two major defects. Firstly, networks with self-attention usually sacrifice internal resolution to balance model performance against complexity, losing fine-grained high-resolution (HR) features. Secondly, even if the optimization focusing on spatial-spectral domain learning (SDL) converges to the ideal solution, there is still a significant visual difference between the reconstructed HSI and the truth. Therefore, we propose a high-resolution dual-domain learning network (HDNet) for HSI reconstruction. On the one hand, the proposed HR spatial-spectral attention module with its efficient feature fusion provides continuous and fine pixel-level features. On the other hand, frequency domain learning (FDL) is introduced for HSI reconstruction to narrow the frequency domain discrepancy. Dynamic FDL supervision forces the model to reconstruct fine-grained frequencies and compensate for excessive smoothing and distortion caused by pixel-level losses. The HR pixel-level attention and frequency-level refinement in our HDNet mutually promote HSI perceptual quality. Extensive quantitative and qualitative evaluation experiments show that our method achieves SOTA performance on simulated and real HSI datasets. Code and models will be released at https://github.com/caiyuanhao1998/MST

Yuanhao Cai, Jing Lin, Haoqian Wang et al.

In coded aperture snapshot spectral compressive imaging (CASSI) systems, hyperspectral image (HSI) reconstruction methods are employed to recover the spatial-spectral signal from a compressed measurement. Among these algorithms, deep unfolding methods demonstrate promising performance but suffer from two issues. Firstly, they do not estimate the degradation patterns and ill-posedness degree from the highly related CASSI to guide the iterative learning. Secondly, they are mainly CNN-based, showing limitations in capturing long-range dependencies. In this paper, we propose a principled Degradation-Aware Unfolding Framework (DAUF) that estimates parameters from the compressed image and physical mask, and then uses these parameters to control each iteration. Moreover, we customize a novel Half-Shuffle Transformer (HST) that simultaneously captures local contents and non-local dependencies. By plugging HST into DAUF, we establish the first Transformer-based deep unfolding method, Degradation-Aware Unfolding Half-Shuffle Transformer (DAUHST), for HSI reconstruction. Experiments show that DAUHST significantly surpasses state-of-the-art methods while requiring cheaper computational and memory costs. Code and models will be released at https://github.com/caiyuanhao1998/MST

Yuanhao Cai, Jing Lin, Xiaowan Hu et al.

Many algorithms have been developed to solve the inverse problem of coded aperture snapshot spectral imaging (CASSI), i.e., recovering the 3D hyperspectral images (HSIs) from a 2D compressive measurement. In recent years, learning-based methods have demonstrated promising performance and dominated the mainstream research direction. However, existing CNN-based methods show limitations in capturing long-range dependencies and non-local self-similarity. Previous Transformer-based methods densely sample tokens, some of which are uninformative, and calculate the multi-head self-attention (MSA) between some tokens that are unrelated in content. This does not fit the spatially sparse nature of HSI signals and limits the model scalability. In this paper, we propose a novel Transformer-based method, coarse-to-fine sparse Transformer (CST), firstly embedding HSI sparsity into deep learning for HSI reconstruction. In particular, CST uses our proposed spectra-aware screening mechanism (SASM) for coarse patch selecting. Then the selected patches are fed into our customized spectra-aggregation hashing multi-head self-attention (SAH-MSA) for fine pixel clustering and self-similarity capturing. Comprehensive experiments show that our CST significantly outperforms state-of-the-art methods while requiring cheaper computational costs. The code and models will be released at https://github.com/caiyuanhao1998/MST

Yuanhao Cai, Xiaowan Hu, Haoqian Wang et al.

Existing deep learning real denoising methods require a large amount of noisy-clean image pairs for supervision. Nonetheless, capturing a real noisy-clean dataset is an unacceptable expensive and cumbersome procedure. To alleviate this problem, this work investigates how to generate realistic noisy images. Firstly, we formulate a simple yet reasonable noise model that treats each real noisy pixel as a random variable. This model splits the noisy image generation problem into two sub-problems: image domain alignment and noise domain alignment. Subsequently, we propose a novel framework, namely Pixel-level Noise-aware Generative Adversarial Network (PNGAN). PNGAN employs a pre-trained real denoiser to map the fake and real noisy images into a nearly noise-free solution space to perform image domain alignment. Simultaneously, PNGAN establishes a pixel-level adversarial training to conduct noise domain alignment. Additionally, for better noise fitting, we present an efficient architecture Simple Multi-scale Network (SMNet) as the generator. Qualitative validation shows that noise generated by PNGAN is highly similar to real noise in terms of intensity and distribution. Quantitative experiments demonstrate that a series of denoisers trained with the generated noisy images achieve state-of-the-art (SOTA) results on four real denoising benchmarks. Part of codes, pre-trained models, and results are available at https://github.com/caiyuanhao1998/PNGAN for comparisons.

Jing Lin, Xiaowan Hu, Yuanhao Cai et al.

How to properly model the inter-frame relation within the video sequence is an important but unsolved challenge for video restoration (VR). In this work, we propose an unsupervised flow-aligned sequence-to-sequence model (S2SVR) to address this problem. On the one hand, the sequence-to-sequence model, which has proven capable of sequence modeling in the field of natural language processing, is explored for the first time in VR. Optimized serialization modeling shows potential in capturing long-range dependencies among frames. On the other hand, we equip the sequence-to-sequence model with an unsupervised optical flow estimator to maximize its potential. The flow estimator is trained with our proposed unsupervised distillation loss, which can alleviate the data discrepancy and inaccurate degraded optical flow issues of previous flow-based methods. With reliable optical flow, we can establish accurate correspondence among multiple frames, narrowing the domain difference between 1D language and 2D misaligned frames and improving the potential of the sequence-to-sequence model. S2SVR shows superior performance in multiple VR tasks, including video deblurring, video super-resolution, and compressed video quality enhancement. Code and models are publicly available at https://github.com/linjing7/VR-Baseline

Haohan Wang, Liang Liu, Wuhao Zhang et al.

Few-shot semantic segmentation aims to learn to segment unseen class objects with the guidance of only a few support images. Most previous methods rely on the pixel-level label of support images. In this paper, we focus on a more challenging setting, in which only the image-level labels are available. We propose a general framework to firstly generate coarse masks with the help of the powerful vision-language model CLIP, and then iteratively and mutually refine the mask predictions of support and query images. Extensive experiments on PASCAL-5i and COCO-20i datasets demonstrate that our method not only outperforms the state-of-the-art weakly supervised approaches by a significant margin, but also achieves comparable or better results to recent supervised methods. Moreover, our method owns an excellent generalization ability for the images in the wild and uncommon classes. Code will be available at https://github.com/Whileherham/IMR-HSNet.

Haohan Wang, Liang Liu, Boshen Zhang et al.

Fully supervised object detection requires training images in which all instances are annotated. This is actually impractical due to the high labor and time costs and the unavoidable missing annotations. As a result, the incomplete annotation in each image could provide misleading supervision and harm the training. Recent works on sparsely annotated object detection alleviate this problem by generating pseudo labels for the missing annotations. Such a mechanism is sensitive to the threshold of the pseudo label score. However, the effective threshold is different in different training stages and among different object detectors. Therefore, the current methods with fixed thresholds have sub-optimal performance, and are difficult to be applied to other detectors. In order to resolve this obstacle, we propose a Calibrated Teacher, of which the confidence estimation of the prediction is well calibrated to match its real precision. In this way, different detectors in different training stages would share a similar distribution of the output confidence, so that multiple detectors could share the same fixed threshold and achieve better performance. Furthermore, we present a simple but effective Focal IoU Weight (FIoU) for the classification loss. FIoU aims at reducing the loss weight of false negative samples caused by the missing annotation, and thus works as the complement of the teacher-student paradigm. Extensive experiments show that our methods set new state-of-the-art under all different sparse settings in COCO. Code will be available at https://github.com/Whileherham/CalibratedTeacher.

Jing Lin, Ailing Zeng, Haoqian Wang et al.

Whole-body mesh recovery aims to estimate the 3D human body, face, and hands parameters from a single image. It is challenging to perform this task with a single network due to resolution issues, i.e., the face and hands are usually located in extremely small regions. Existing works usually detect hands and faces, enlarge their resolution to feed in a specific network to predict the parameter, and finally fuse the results. While this copy-paste pipeline can capture the fine-grained details of the face and hands, the connections between different parts cannot be easily recovered in late fusion, leading to implausible 3D rotation and unnatural pose. In this work, we propose a one-stage pipeline for expressive whole-body mesh recovery, named OSX, without separate networks for each part. Specifically, we design a Component Aware Transformer (CAT) composed of a global body encoder and a local face/hand decoder. The encoder predicts the body parameters and provides a high-quality feature map for the decoder, which performs a feature-level upsample-crop scheme to extract high-resolution part-specific features and adopt keypoint-guided deformable attention to estimate hand and face precisely. The whole pipeline is simple yet effective without any manual post-processing and naturally avoids implausible prediction. Comprehensive experiments demonstrate the effectiveness of OSX. Lastly, we build a large-scale Upper-Body dataset (UBody) with high-quality 2D and 3D whole-body annotations. It contains persons with partially visible bodies in diverse real-life scenarios to bridge the gap between the basic task and downstream applications.

Zhuoling Li, Chuanrui Zhang, Wei-Chiu Ma et al.

In recent years, transformer-based detectors have demonstrated remarkable performance in 2D visual perception tasks. However, their performance in multi-view 3D object detection remains inferior to the state-of-the-art (SOTA) of convolutional neural network based detectors. In this work, we investigate this issue from the perspective of bird's-eye-view (BEV) feature generation. Specifically, we examine the BEV feature generation method employed by the transformer-based SOTA, BEVFormer, and identify its two limitations: (i) it only generates attention weights from BEV, which precludes the use of lidar points for supervision, and (ii) it aggregates camera view features to the BEV through deformable sampling, which only selects a small subset of features and fails to exploit all information. To overcome these limitations, we propose a novel BEV feature generation method, dual-view attention, which generates attention weights from both the BEV and camera view. This method encodes all camera features into the BEV feature. By combining dual-view attention with the BEVFormer architecture, we build a new detector named VoxelFormer. Extensive experiments are conducted on the nuScenes benchmark to verify the superiority of dual-view attention and VoxelForer. We observe that even only adopting 3 encoders and 1 historical frame during training, VoxelFormer still outperforms BEVFormer significantly. When trained in the same setting, VoxelFormer can surpass BEVFormer by 4.9% NDS point. Code is available at: https://github.com/Lizhuoling/VoxelFormer-public.git.

Hao Bian, Zhuchen Shao, Yang Chen et al.

With the development of computational pathology, deep learning methods for Gleason grading through whole slide images (WSIs) have excellent prospects. Since the size of WSIs is extremely large, the image label usually contains only slide-level label or limited pixel-level labels. The current mainstream approach adopts multi-instance learning to predict Gleason grades. However, some methods only considering the slide-level label ignore the limited pixel-level labels containing rich local information. Furthermore, the method of additionally considering the pixel-level labels ignores the inaccuracy of pixel-level labels. To address these problems, we propose a mixed supervision Transformer based on the multiple instance learning framework. The model utilizes both slide-level label and instance-level labels to achieve more accurate Gleason grading at the slide level. The impact of inaccurate instance-level labels is further reduced by introducing an efficient random masking strategy in the mixed supervision training process. We achieve the state-of-the-art performance on the SICAPv2 dataset, and the visual analysis shows the accurate prediction results of instance level. The source code is available at https://github.com/bianhao123/Mixed_supervision.

Zhuoling Li, Zhan Qu, Yang Zhou et al.

As an inherently ill-posed problem, depth estimation from single images is the most challenging part of monocular 3D object detection (M3OD). Many existing methods rely on preconceived assumptions to bridge the missing spatial information in monocular images, and predict a sole depth value for every object of interest. However, these assumptions do not always hold in practical applications. To tackle this problem, we propose a depth solving system that fully explores the visual clues from the subtasks in M3OD and generates multiple estimations for the depth of each target. Since the depth estimations rely on different assumptions in essence, they present diverse distributions. Even if some assumptions collapse, the estimations established on the remaining assumptions are still reliable. In addition, we develop a depth selection and combination strategy. This strategy is able to remove abnormal estimations caused by collapsed assumptions, and adaptively combine the remaining estimations into a single one. In this way, our depth solving system becomes more precise and robust. Exploiting the clues from multiple subtasks of M3OD and without introducing any extra information, our method surpasses the current best method by more than 20% relatively on the Moderate level of test split in the KITTI 3D object detection benchmark, while still maintaining real-time efficiency.

Jing Lin, Ailing Zeng, Shunlin Lu et al.

In this paper, we present Motion-X, a large-scale 3D expressive whole-body motion dataset. Existing motion datasets predominantly contain body-only poses, lacking facial expressions, hand gestures, and fine-grained pose descriptions. Moreover, they are primarily collected from limited laboratory scenes with textual descriptions manually labeled, which greatly limits their scalability. To overcome these limitations, we develop a whole-body motion and text annotation pipeline, which can automatically annotate motion from either single- or multi-view videos and provide comprehensive semantic labels for each video and fine-grained whole-body pose descriptions for each frame. This pipeline is of high precision, cost-effective, and scalable for further research. Based on it, we construct Motion-X, which comprises 15.6M precise 3D whole-body pose annotations (i.e., SMPL-X) covering 81.1K motion sequences from massive scenes. Besides, Motion-X provides 15.6M frame-level whole-body pose descriptions and 81.1K sequence-level semantic labels. Comprehensive experiments demonstrate the accuracy of the annotation pipeline and the significant benefit of Motion-X in enhancing expressive, diverse, and natural motion generation, as well as 3D whole-body human mesh recovery.

Ruichen Zheng, Peng Li, Haoqian Wang et al.

Detailed 3D reconstruction and photo-realistic relighting of digital humans are essential for various applications. To this end, we propose a novel sparse-view 3d human reconstruction framework that closely incorporates the occupancy field and albedo field with an additional visibility field--it not only resolves occlusion ambiguity in multiview feature aggregation, but can also be used to evaluate light attenuation for self-shadowed relighting. To enhance its training viability and efficiency, we discretize visibility onto a fixed set of sample directions and supply it with coupled geometric 3D depth feature and local 2D image feature. We further propose a novel rendering-inspired loss, namely TransferLoss, to implicitly enforce the alignment between visibility and occupancy field, enabling end-to-end joint training. Results and extensive experiments demonstrate the effectiveness of the proposed method, as it surpasses state-of-the-art in terms of reconstruction accuracy while achieving comparably accurate relighting to ray-traced ground truth.

Shiqi Xu, Wenhui Liu, Xi Yang et al.

Fast noninvasive probing of spatially varying decorrelating events, such as cerebral blood flow beneath the human skull, is an essential task in various scientific and clinical settings. One of the primary optical techniques used is diffuse correlation spectroscopy (DCS), whose classical implementation uses a single or few single-photon detectors, resulting in poor spatial localization accuracy and relatively low temporal resolution. Here, we propose a technique termed Classifying Rapid decorrelation Events via Parallelized single photon dEtection (CREPE)}, a new form of DCS that can probe and classify different decorrelating movements hidden underneath turbid volume with high sensitivity using parallelized speckle detection from a $32\times32$ pixel SPAD array. We evaluate our setup by classifying different spatiotemporal-decorrelating patterns hidden beneath a 5mm tissue-like phantom made with rapidly decorrelating dynamic scattering media. Twelve multi-mode fibers are used to collect scattered light from different positions on the surface of the tissue phantom. To validate our setup, we generate perturbed decorrelation patterns by both a digital micromirror device (DMD) modulated at multi-kilo-hertz rates, as well as a vessel phantom containing flowing fluid. Along with a deep contrastive learning algorithm that outperforms classic unsupervised learning methods, we demonstrate our approach can accurately detect and classify different transient decorrelation events (happening in 0.1-0.4s) underneath turbid scattering media, without any data labeling. This has the potential to be applied to noninvasively monitor deep tissue motion patterns, for example identifying normal or abnormal cerebral blood flow events, at multi-Hertz rates within a compact and static detection probe.

Luyuan Zhang, Siyuan Li, Zedong Wang et al.

Most visual tokenizers for image generation are bifurcated into two families with complementary limitations: continuous VAEs offer high-fidelity reconstruction but suffer from dense, entangled latents that are poorly suited for semantic control, whereas discrete VQ-based models enable autoregressive generation yet struggle with gradient sparsity, unstable training, and codebook collapse. In this work, we introduce MergeTok, a unified tokenizer that jointly optimizes continuous (VAE) and discrete (VQ) tokenizers within a encoder-decoder architecture, leveraging token merging techniques as a semantic bridge. By clustering similar tokens during encoding, MergeTok establishes a structural prior that provides dual supervision signals: (i) it imposes merged-token semantic alignment in the VAE branch, regularizing its latent space toward disentangled, semantic-aware representations; (ii) it derives group-wise constraints, promoting intra-group diversity and inter-group exclusivity that stabilize VQ training. MergeTok shows competitive reconstruction and generation performance on ImageNet-256, with substantially lower rFID than strong VAE and VQ models under matched token budgets, while producing semantically-organized token representations compatible with both autoregressive and diffusion generators. This shows that a single architecture can endow visual tokenizers with robust semantic organization and generator-friendly discreteness.

Yuxiao Yang, Xiao-Xiao Long, Zhiyang Dou et al.

In this work, we introduce \textbf{Wonder3D++}, a novel method for efficiently generating high-fidelity textured meshes from single-view images. Recent methods based on Score Distillation Sampling (SDS) have shown the potential to recover 3D geometry from 2D diffusion priors, but they typically suffer from time-consuming per-shape optimization and inconsistent geometry. In contrast, certain works directly produce 3D information via fast network inferences, but their results are often of low quality and lack geometric details. To holistically improve the quality, consistency, and efficiency of single-view reconstruction tasks, we propose a cross-domain diffusion model that generates multi-view normal maps and the corresponding color images. To ensure the consistency of generation, we employ a multi-view cross-domain attention mechanism that facilitates information exchange across views and modalities. Lastly, we introduce a cascaded 3D mesh extraction algorithm that drives high-quality surfaces from the multi-view 2D representations in only about $3$ minute in a coarse-to-fine manner. Our extensive evaluations demonstrate that our method achieves high-quality reconstruction results, robust generalization, and good efficiency compared to prior works. Code available at https://github.com/xxlong0/Wonder3D/tree/Wonder3D_Plus.

Zhuoling Li, Chunrui Han, Zheng Ge et al.

Efficiency is quite important for 3D lane detection due to practical deployment demand. In this work, we propose a simple, fast, and end-to-end detector that still maintains high detection precision. Specifically, we devise a set of fully convolutional heads based on row-wise classification. In contrast to previous counterparts, ours supports recognizing both vertical and horizontal lanes. Besides, our method is the first one to perform row-wise classification in bird-eye-view. In the heads, we split feature into multiple groups and every group of feature corresponds to a lane instance. During training, the predictions are associated with lane labels using the proposed single-win one-to-one matching to compute loss, and no post-processing operation is demanded for inference. In this way, our proposed fully convolutional detector, GroupLane, realizes end-to-end detection like DETR. Evaluated on 3 real world 3D lane benchmarks, OpenLane, Once-3DLanes, and OpenLane-Huawei, GroupLane adopting ConvNext-Base as the backbone outperforms the published state-of-the-art PersFormer by 13.6% F1 score in the OpenLane validation set. Besides, GroupLane with ResNet18 still surpasses PersFormer by 4.9% F1 score, while the inference speed is nearly 7x faster and the FLOPs is only 13.3% of it.

Xiaowan Hu, Yiyi Chen, Yan Li et al.

With the rapid expansion of e-commerce, more consumers have become accustomed to making purchases via livestreaming. Accurately identifying the products being sold by salespeople, i.e., livestreaming product retrieval (LPR), poses a fundamental and daunting challenge. The LPR task encompasses three primary dilemmas in real-world scenarios: 1) the recognition of intended products from distractor products present in the background; 2) the video-image heterogeneity that the appearance of products showcased in live streams often deviates substantially from standardized product images in stores; 3) there are numerous confusing products with subtle visual nuances in the shop. To tackle these challenges, we propose the Spatiotemporal Graphing Multi-modal Network (SGMN). First, we employ a text-guided attention mechanism that leverages the spoken content of salespeople to guide the model to focus toward intended products, emphasizing their salience over cluttered background products. Second, a long-range spatiotemporal graph network is further designed to achieve both instance-level interaction and frame-level matching, solving the misalignment caused by video-image heterogeneity. Third, we propose a multi-modal hard example mining, assisting the model in distinguishing highly similar products with fine-grained features across the video-image-text domain. Through extensive quantitative and qualitative experiments, we demonstrate the superior performance of our proposed SGMN model, surpassing the state-of-the-art methods by a substantial margin. The code is available at https://github.com/Huxiaowan/SGMN.

Zhuohong Chen, Zhengxian Wu, Zirui Liao et al.

Vision-centric retrieval for VQA requires retrieving images to supply missing visual cues and integrating them into the reasoning process. However, selecting the right images and integrating them effectively into the model's reasoning remains challenging.To address this challenge, we propose R3G, a modular Reasoning-Retrieval-Reranking framework.It first produces a brief reasoning plan that specifies the required visual cues, then adopts a two-stage strategy, with coarse retrieval followed by fine-grained reranking, to select evidence images.On MRAG-Bench, R3G improves accuracy across six MLLM backbones and nine sub-scenarios, achieving state-of-the-art overall performance. Ablations show that sufficiency-aware reranking and reasoning steps are complementary, helping the model both choose the right images and use them well. We release code and data at https://github.com/czh24/R3G.

Zhengxian Wu, Kai Shi, Chuanrui Zhang et al.

Recent progress in multimodal large language models has led to strong performance on reasoning tasks, but these improvements largely rely on high-quality annotated data or teacher-model distillation, both of which are costly and difficult to scale. To address this, we propose an unsupervised self-evolution training framework for multimodal reasoning that achieves stable performance improvements without using human-annotated answers or external reward models. For each input, we sample multiple reasoning trajectories and jointly model their within group structure. We use the Actor's self-consistency signal as a training prior, and introduce a bounded Judge based modulation to continuously reweight trajectories of different quality. We further model the modulated scores as a group level distribution and convert absolute scores into relative advantages within each group, enabling more robust policy updates. Trained with Group Relative Policy Optimization (GRPO) on unlabeled data, our method consistently improves reasoning performance and generalization on five mathematical reasoning benchmarks, offering a scalable path toward self-evolving multimodal models. The code are available at https://github.com/OPPO-Mente-Lab/LLM-Self-Judge.

Zhe Huang, Hao Wen, Aiming Hao et al.

Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding. However, they suffer from a critical vulnerability: an over-reliance on language priors, which can lead to visual ungrounded hallucinations, especially when processing counterfactual videos that defy common sense. This limitation, stemming from the intrinsic data imbalance between text and video, is challenging to address due to the substantial cost of collecting and annotating counterfactual data. To address this, we introduce DualityForge, a novel counterfactual data synthesis framework that employs controllable, diffusion-based video editing to transform real-world videos into counterfactual scenarios. By embedding structured contextual information into the video editing and QA generation processes, the framework automatically produces high-quality QA pairs together with original-edited video pairs for contrastive training. Based on this, we build DualityVidQA, a large-scale video dataset designed to reduce MLLM hallucinations. In addition, to fully exploit the contrastive nature of our paired data, we propose Duality-Normalized Advantage Training (DNA-Train), a two-stage SFT-RL training regime where the RL phase applies pair-wise $\ell_1$ advantage normalization, thereby enabling a more stable and efficient policy optimization. Experiments on DualityVidQA-Test demonstrate that our method substantially reduces model hallucinations on counterfactual videos, yielding a relative improvement of 24.0% over the Qwen2.5-VL-7B baseline. Moreover, our approach achieves significant gains across both hallucination and general-purpose benchmarks, indicating strong generalization capability. We will open-source our dataset and code.

Chendong Zhao, Jianzong Wang, Wen qi Wei et al.

The Transformer architecture model, based on self-attention and multi-head attention, has achieved remarkable success in offline end-to-end Automatic Speech Recognition (ASR). However, self-attention and multi-head attention cannot be easily applied for streaming or online ASR. For self-attention in Transformer ASR, the softmax normalization function-based attention mechanism makes it impossible to highlight important speech information. For multi-head attention in Transformer ASR, it is not easy to model monotonic alignments in different heads. To overcome these two limits, we integrate sparse attention and monotonic attention into Transformer-based ASR. The sparse mechanism introduces a learned sparsity scheme to enable each self-attention structure to fit the corresponding head better. The monotonic attention deploys regularization to prune redundant heads for the multi-head attention structure. The experiments show that our method can effectively improve the attention mechanism on widely used benchmarks of speech recognition.

Chuanrui Zhang, Yingshuang Zou, Zhuoling Li et al.

Compared with previous 3D reconstruction methods like Nerf, recent Generalizable 3D Gaussian Splatting (G-3DGS) methods demonstrate impressive efficiency even in the sparse-view setting. However, the promising reconstruction performance of existing G-3DGS methods relies heavily on accurate multi-view feature matching, which is quite challenging. Especially for the scenes that have many non-overlapping areas between various views and contain numerous similar regions, the matching performance of existing methods is poor and the reconstruction precision is limited. To address this problem, we develop a strategy that utilizes a predicted depth confidence map to guide accurate local feature matching. In addition, we propose to utilize the knowledge of existing monocular depth estimation models as prior to boost the depth estimation precision in non-overlapping areas between views. Combining the proposed strategies, we present a novel G-3DGS method named TranSplat, which obtains the best performance on both the RealEstate10K and ACID benchmarks while maintaining competitive speed and presenting strong cross-dataset generalization ability. Our code, and demos will be available at: https://xingyoujun.github.io/transplat.

Minghan Qin, Yifan Liu, Yuelang Xu et al.

One crucial aspect of 3D head avatar reconstruction lies in the details of facial expressions. Although recent NeRF-based photo-realistic 3D head avatar methods achieve high-quality avatar rendering, they still encounter challenges retaining intricate facial expression details because they overlook the potential of specific expression variations at different spatial positions when conditioning the radiance field. Motivated by this observation, we introduce a novel Spatially-Varying Expression (SVE) conditioning. The SVE can be obtained by a simple MLP-based generation network, encompassing both spatial positional features and global expression information. Benefiting from rich and diverse information of the SVE at different positions, the proposed SVE-conditioned neural radiance field can deal with intricate facial expressions and achieve realistic rendering and geometry details of high-fidelity 3D head avatars. Additionally, to further elevate the geometric and rendering quality, we introduce a new coarse-to-fine training strategy, including a geometry initialization strategy at the coarse stage and an adaptive importance sampling strategy at the fine stage. Extensive experiments indicate that our method outperforms other state-of-the-art (SOTA) methods in rendering and geometry quality on mobile phone-collected and public datasets.

Yifan Liu, Youbao Tang, Ning Zhang et al.

Temporal action localization (TAL) aims to detect the boundary and identify the class of each action instance in a long untrimmed video. Current approaches treat video frames homogeneously, and tend to give background and key objects excessive attention. This limits their sensitivity to localize action boundaries. To this end, we propose a prior-enhanced temporal action localization method (PETAL), which only takes in RGB input and incorporates action subjects as priors. This proposal leverages action subjects' information with a plug-and-play subject-aware spatial attention module (SA-SAM) to generate an aggregated and subject-prioritized representation. Experimental results on THUMOS-14 and ActivityNet-1.3 datasets demonstrate that the proposed PETAL achieves competitive performance using only RGB features, e.g., boosting mAP by 2.41% or 0.25% over the state-of-the-art approach that uses RGB features or with additional optical flow features on the THUMOS-14 dataset.

Yang Liu, Xiang Huang, Minghan Qin et al.

Neural radiance fields are capable of reconstructing high-quality drivable human avatars but are expensive to train and render and not suitable for multi-human scenes with complex shadows. To reduce consumption, we propose Animatable 3D Gaussian, which learns human avatars from input images and poses. We extend 3D Gaussians to dynamic human scenes by modeling a set of skinned 3D Gaussians and a corresponding skeleton in canonical space and deforming 3D Gaussians to posed space according to the input poses. We introduce a multi-head hash encoder for pose-dependent shape and appearance and a time-dependent ambient occlusion module to achieve high-quality reconstructions in scenes containing complex motions and dynamic shadows. On both novel view synthesis and novel pose synthesis tasks, our method achieves higher reconstruction quality than InstantAvatar with less training time (1/60), less GPU memory (1/4), and faster rendering speed (7x). Our method can be easily extended to multi-human scenes and achieve comparable novel view synthesis results on a scene with ten people in only 25 seconds of training.

Peng Li, Jiayin Zhao, Jingyao Wu et al.

Light field disparity estimation is an essential task in computer vision with various applications. Although supervised learning-based methods have achieved both higher accuracy and efficiency than traditional optimization-based methods, the dependency on ground-truth disparity for training limits the overall generalization performance not to say for real-world scenarios where the ground-truth disparity is hard to capture. In this paper, we argue that unsupervised methods can achieve comparable accuracy, but, more importantly, much higher generalization capacity and efficiency than supervised methods. Specifically, we present the Occlusion Pattern Aware Loss, named OPAL, which successfully extracts and encodes the general occlusion patterns inherent in the light field for loss calculation. OPAL enables: i) accurate and robust estimation by effectively handling occlusions without using any ground-truth information for training and ii) much efficient performance by significantly reducing the network parameters required for accurate inference. Besides, a transformer-based network and a refinement module are proposed for achieving even more accurate results. Extensive experiments demonstrate our method not only significantly improves the accuracy compared with the SOTA unsupervised methods, but also possesses strong generalization capacity, even for real-world data, compared with supervised methods. Our code will be made publicly available.

Chendong Zhao, Jianzong Wang, Xiaoyang Qu et al.

Unsupervised representation learning for speech audios attained impressive performances for speech recognition tasks, particularly when annotated speech is limited. However, the unsupervised paradigm needs to be carefully designed and little is known about what properties these representations acquire. There is no guarantee that the model learns meaningful representations for valuable information for recognition. Moreover, the adaptation ability of the learned representations to other domains still needs to be estimated. In this work, we explore learning domain-invariant representations via a direct mapping of speech representations to their corresponding high-level linguistic informations. Results prove that the learned latents not only capture the articulatory feature of each phoneme but also enhance the adaptation ability, outperforming the baseline largely on accented benchmarks.

Zhuoling Li, Chuanrui Zhang, En Yu et al.

The labels of monocular 3D object detection (M3OD) are expensive to obtain. Meanwhile, there usually exists numerous unlabeled data in practical applications, and pre-training is an efficient way of exploiting the knowledge in unlabeled data. However, the pre-training paradigm for M3OD is hardly studied. We aim to bridge this gap in this work. To this end, we first draw two observations: (1) The guideline of devising pre-training tasks is imitating the representation of the target task. (2) Combining depth estimation and 2D object detection is a promising M3OD pre-training baseline. Afterwards, following the guideline, we propose several strategies to further improve this baseline, which mainly include target guided semi-dense depth estimation, keypoint-aware 2D object detection, and class-level loss adjustment. Combining all the developed techniques, the obtained pre-training framework produces pre-trained backbones that improve M3OD performance significantly on both the KITTI-3D and nuScenes benchmarks. For example, by applying a DLA34 backbone to a naive center-based M3OD detector, the moderate ${\rm AP}_{3D}70$ score of Car on the KITTI-3D testing set is boosted by 18.71\% and the NDS score on the nuScenes validation set is improved by 40.41\% relatively.

Zhuchen Shao, Liuxi Dai, Yifeng Wang et al.

Multiple Instance Learning (MIL), a powerful strategy for weakly supervised learning, is able to perform various prediction tasks on gigapixel Whole Slide Images (WSIs). However, the tens of thousands of patches in WSIs usually incur a vast computational burden for image augmentation, limiting the MIL model's improvement in performance. Currently, the feature augmentation-based MIL framework is a promising solution, while existing methods such as Mixup often produce unrealistic features. To explore a more efficient and practical augmentation method, we introduce the Diffusion Model (DM) into MIL for the first time and propose a feature augmentation framework called AugDiff. Specifically, we employ the generation diversity of DM to improve the quality of feature augmentation and the step-by-step generation property to control the retention of semantic information. We conduct extensive experiments over three distinct cancer datasets, two different feature extractors, and three prevalent MIL algorithms to evaluate the performance of AugDiff. Ablation study and visualization further verify the effectiveness. Moreover, we highlight AugDiff's higher-quality augmented feature over image augmentation and its superiority over self-supervised learning. The generalization over external datasets indicates its broader applications.

Chuanrui Zhang, Yonggen Ling, Minglei Lu et al.

We study the 3D object understanding task for manipulating everyday objects with different material properties (diffuse, specular, transparent and mixed). Existing monocular and RGB-D methods suffer from scale ambiguity due to missing or imprecise depth measurements. We present CODERS, a one-stage approach for Category-level Object Detection, pose Estimation and Reconstruction from Stereo images. The base of our pipeline is an implicit stereo matching module that combines stereo image features with 3D position information. Concatenating this presented module and the following transform-decoder architecture leads to end-to-end learning of multiple tasks required by robot manipulation. Our approach significantly outperforms all competing methods in the public TOD dataset. Furthermore, trained on simulated data, CODERS generalize well to unseen category-level object instances in real-world robot manipulation experiments. Our dataset, code, and demos will be available on our project page.

Yunyao Yu, Zhengxian Wu, Zhuohong Chen et al.

In the unsupervised self-evolution of Multimodal Large Language Models, the quality of feedback signals during post-training is pivotal for stable and effective learning. However, existing self-evolution methods predominantly rely on majority voting to select the most frequent output as the pseudo-golden answer, which may stem from the model's intrinsic biases rather than guaranteeing the objective correctness of the reasoning paths. To counteract the degradation, we propose \textbf{C}ontinuous \textbf{S}oftened \textbf{R}etracing re\textbf{S}ampling (\textbf{CSRS}) in MLLM self-evolution. Specifically, we introduce a Retracing Re-inference Mechanism (\textbf{RRM}) that the model re-inferences from anchor points to expand the exploration of long-tail reasoning paths. Simultaneously, we propose Softened Frequency Reward (\textbf{SFR}), which replaces binary rewards with continuous signals, calibrating reward based on the answers' frequency across sampled reasoning sets. Furthermore, incorporated with Visual Semantic Perturbation (\textbf{VSP}), CSRS ensures the model prioritizes mathematical logic over visual superficiality. Experimental results demonstrate that CSRS significantly enhances the reasoning performance of Qwen2.5-VL-7B on benchmarks such as MathVision. We achieve state-of-the-art (SOTA) results in unsupervised self-evolution on geometric tasks. Our code is avaible at https://github.com/yyy195/CSRS.

Xiangwen Deng, Yingshuang Zou, Yuanhao Cai et al.

Style transfer of 3D faces has gained more and more attention. However, previous methods mainly use images of artistic faces for style transfer while ignoring arbitrary style images such as abstract paintings. To solve this problem, we propose a novel method, namely Face-guided Dual Style Transfer (FDST). To begin with, FDST employs a 3D decoupling module to separate facial geometry and texture. Then we propose a style fusion strategy for facial geometry. Subsequently, we design an optimization-based DDSG mechanism for textures that can guide the style transfer by two style images. Besides the normal style image input, DDSG can utilize the original face input as another style input as the face prior. By this means, high-quality face arbitrary style transfer results can be obtained. Furthermore, FDST can be applied in many downstream tasks, including region-controllable style transfer, high-fidelity face texture reconstruction, large-pose face reconstruction, and artistic face reconstruction. Comprehensive quantitative and qualitative results show that our method can achieve comparable performance. All source codes and pre-trained weights will be released to the public.

Huaqiu Li, Xiaowan Hu, Haoqian Wang

Real-world low-light images often suffer from complex degradations such as local overexposure, low brightness, noise, and uneven illumination. Supervised methods tend to overfit to specific scenarios, while unsupervised methods, though better at generalization, struggle to model these degradations due to the lack of reference images. To address this issue, we propose an interpretable, zero-reference joint denoising and low-light enhancement framework tailored for real-world scenarios. Our method derives a training strategy based on paired sub-images with varying illumination and noise levels, grounded in physical imaging principles and retinex theory. Additionally, we leverage the Discrete Cosine Transform (DCT) to perform frequency domain decomposition in the sRGB space, and introduce an implicit-guided hybrid representation strategy that effectively separates intricate compounded degradations. In the backbone network design, we develop retinal decomposition network guided by implicit degradation representation mechanisms. Extensive experiments demonstrate the superiority of our method. Code will be available at https://github.com/huaqlili/unsupervised-light-enhance-ICLR2025.

Minghan Qin, Wanhua Li, Jiawei Zhou et al. · tsinghua

Humans live in a 3D world and commonly use natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space, thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experimental results show that LangSplat significantly outperforms the previous state-of-the-art method LERF by a large margin. Notably, LangSplat is extremely efficient, achieving a 199 $\times$ speedup compared to LERF at the resolution of 1440 $\times$ 1080. We strongly recommend readers to check out our video results at https://langsplat.github.io/

Huaqiu Li, Yong Wang, Tongwen Huang et al.

Unified image restoration is a significantly challenging task in low-level vision. Existing methods either make tailored designs for specific tasks, limiting their generalizability across various types of degradation, or rely on training with paired datasets, thereby suffering from closed-set constraints. To address these issues, we propose a novel, dataset-free, and unified approach through recurrent posterior sampling utilizing a pretrained latent diffusion model. Our method incorporates the multimodal understanding model to provide sematic priors for the generative model under a task-blind condition. Furthermore, it utilizes a lightweight module to align the degraded input with the generated preference of the diffusion model, and employs recurrent refinement for posterior sampling. Extensive experiments demonstrate that our method outperforms state-of-the-art methods, validating its effectiveness and robustness. Our code and data are available at https://github.com/AMAP-ML/LD-RPS.

Huaqiu Li, Wang Zhang, Xiaowan Hu et al.

Many studies have concentrated on constructing supervised models utilizing paired datasets for image denoising, which proves to be expensive and time-consuming. Current self-supervised and unsupervised approaches typically rely on blind-spot networks or sub-image pairs sampling, resulting in pixel information loss and destruction of detailed structural information, thereby significantly constraining the efficacy of such methods. In this paper, we introduce Prompt-SID, a prompt-learning-based single image denoising framework that emphasizes preserving of structural details. This approach is trained in a self-supervised manner using downsampled image pairs. It captures original-scale image information through structural encoding and integrates this prompt into the denoiser. To achieve this, we propose a structural representation generation model based on the latent diffusion process and design a structural attention module within the transformer-based denoiser architecture to decode the prompt. Additionally, we introduce a scale replay training mechanism, which effectively mitigates the scale gap from images of different resolutions. We conduct comprehensive experiments on synthetic, real-world, and fluorescence imaging datasets, showcasing the remarkable effectiveness of Prompt-SID. Our code will be released at https://github.com/huaqlili/Prompt-SID.

Qinwei Lin, Xiaopeng Sun, Yu Gao et al.

Diffusion models have recently achieved outstanding results in the field of image super-resolution. These methods typically inject low-resolution (LR) images via ControlNet.In this paper, we first explore the temporal dynamics of information infusion through ControlNet, revealing that the input from LR images predominantly influences the initial stages of the denoising process. Leveraging this insight, we introduce a novel timestep-aware diffusion model that adaptively integrates features from both ControlNet and the pre-trained Stable Diffusion (SD). Our method enhances the transmission of LR information in the early stages of diffusion to guarantee image fidelity and stimulates the generation ability of the SD model itself more in the later stages to enhance the detail of generated images. To train this method, we propose a timestep-aware training strategy that adopts distinct losses at varying timesteps and acts on disparate modules. Experiments on benchmark datasets demonstrate the effectiveness of our method. Code: https://github.com/SleepyLin/TASR

Yuxiao Yang, Hualian Sheng, Sijia Cai et al.

Video generation models have advanced significantly, yet they still struggle to synthesize complex human movements due to the high degrees of freedom in human articulation. This limitation stems from the intrinsic constraints of pixel-only training objectives, which inherently bias models toward appearance fidelity at the expense of learning underlying kinematic principles. To address this, we introduce EchoMotion, a framework designed to model the joint distribution of appearance and human motion, thereby improving the quality of complex human action video generation. EchoMotion extends the DiT (Diffusion Transformer) framework with a dual-branch architecture that jointly processes tokens concatenated from different modalities. Furthermore, we propose MVS-RoPE (Motion-Video Syncronized RoPE), which offers unified 3D positional encoding for both video and motion tokens. By providing a synchronized coordinate system for the dual-modal latent sequence, MVS-RoPE establishes an inductive bias that fosters temporal alignment between the two modalities. We also propose a Motion-Video Two-Stage Training Strategy. This strategy enables the model to perform both the joint generation of complex human action videos and their corresponding motion sequences, as well as versatile cross-modal conditional generation tasks. To facilitate the training of a model with these capabilities, we construct HuMoVe, a large-scale dataset of approximately 80,000 high-quality, human-centric video-motion pairs. Our findings reveal that explicitly representing human motion is complementary to appearance, significantly boosting the coherence and plausibility of human-centric video generation.

Bingquan Dai, Li Ray Luo, Qihong Tang et al.

Reconstructing 3D objects into editable programs is pivotal for applications like reverse engineering and shape editing. However, existing methods often rely on limited domain-specific languages (DSLs) and small-scale datasets, restricting their ability to model complex geometries and structures. To address these challenges, we introduce MeshCoder, a novel framework that reconstructs complex 3D objects from point clouds into editable Blender Python scripts. We develop a comprehensive set of expressive Blender Python APIs capable of synthesizing intricate geometries. Leveraging these APIs, we construct a large-scale paired object-code dataset, where the code for each object is decomposed into distinct semantic parts. Subsequently, we train a multimodal large language model (LLM) that translates 3D point cloud into executable Blender Python scripts. Our approach not only achieves superior performance in shape-to-code reconstruction tasks but also facilitates intuitive geometric and topological editing through convenient code modifications. Furthermore, our code-based representation enhances the reasoning capabilities of LLMs in 3D shape understanding tasks. Together, these contributions establish MeshCoder as a powerful and flexible solution for programmatic 3D shape reconstruction and understanding. The project homepage is available at \href{https://daibingquan.github.io/MeshCoder}{this link}.

Zikang Chen, Tao Jiang, Xiaowan Hu et al.

Self-supervised video denoising aims to remove noise from videos without relying on ground truth data, leveraging the video itself to recover clean frames. Existing methods often rely on simplistic feature stacking or apply optical flow without thorough analysis. This results in suboptimal utilization of both inter-frame and intra-frame information, and it also neglects the potential of optical flow alignment under self-supervised conditions, leading to biased and insufficient denoising outcomes. To this end, we first explore the practicality of optical flow in the self-supervised setting and introduce a SpatioTemporal Blind-spot Network (STBN) for global frame feature utilization. In the temporal domain, we utilize bidirectional blind-spot feature propagation through the proposed blind-spot alignment block to ensure accurate temporal alignment and effectively capture long-range dependencies. In the spatial domain, we introduce the spatial receptive field expansion module, which enhances the receptive field and improves global perception capabilities. Additionally, to reduce the sensitivity of optical flow estimation to noise, we propose an unsupervised optical flow distillation mechanism that refines fine-grained inter-frame interactions during optical flow alignment. Our method demonstrates superior performance across both synthetic and real-world video denoising datasets. The source code is publicly available at https://github.com/ZKCCZ/STBN.

Yaoyu Su, Shaohui Wang, Haoqian Wang

In this paper, we present the decomposed triplane-hash neural radiance fields (DT-NeRF), a framework that significantly improves the photorealistic rendering of talking faces and achieves state-of-the-art results on key evaluation datasets. Our architecture decomposes the facial region into two specialized triplanes: one specialized for representing the mouth, and the other for the broader facial features. We introduce audio features as residual terms and integrate them as query vectors into our model through an audio-mouth-face transformer. Additionally, our method leverages the capabilities of Neural Radiance Fields (NeRF) to enrich the volumetric representation of the entire face through additive volumetric rendering techniques. Comprehensive experimental evaluations corroborate the effectiveness and superiority of our proposed approach.

Yuanhao Cai, Yuxin Zheng, Jing Lin et al.

Existing deep learning models for hyperspectral image (HSI) reconstruction achieve good performance but require powerful hardwares with enormous memory and computational resources. Consequently, these methods can hardly be deployed on resource-limited mobile devices. In this paper, we propose a novel method, Binarized Spectral-Redistribution Network (BiSRNet), for efficient and practical HSI restoration from compressed measurement in snapshot compressive imaging (SCI) systems. Firstly, we redesign a compact and easy-to-deploy base model to be binarized. Then we present the basic unit, Binarized Spectral-Redistribution Convolution (BiSR-Conv). BiSR-Conv can adaptively redistribute the HSI representations before binarizing activation and uses a scalable hyperbolic tangent function to closer approximate the Sign function in backpropagation. Based on our BiSR-Conv, we customize four binarized convolutional modules to address the dimension mismatch and propagate full-precision information throughout the whole network. Finally, our BiSRNet is derived by using the proposed techniques to binarize the base model. Comprehensive quantitative and qualitative experiments manifest that our proposed BiSRNet outperforms state-of-the-art binarization methods and achieves comparable performance with full-precision algorithms. Code and models are publicly available at https://github.com/caiyuanhao1998/BiSCI and https://github.com/caiyuanhao1998/MST

Jing Lin, Yuanhao Cai, Xiaowan Hu et al.

Exploiting similar and sharper scene patches in spatio-temporal neighborhoods is critical for video deblurring. However, CNN-based methods show limitations in capturing long-range dependencies and modeling non-local self-similarity. In this paper, we propose a novel framework, Flow-Guided Sparse Transformer (FGST), for video deblurring. In FGST, we customize a self-attention module, Flow-Guided Sparse Window-based Multi-head Self-Attention (FGSW-MSA). For each $query$ element on the blurry reference frame, FGSW-MSA enjoys the guidance of the estimated optical flow to globally sample spatially sparse yet highly related $key$ elements corresponding to the same scene patch in neighboring frames. Besides, we present a Recurrent Embedding (RE) mechanism to transfer information from past frames and strengthen long-range temporal dependencies. Comprehensive experiments demonstrate that our proposed FGST outperforms state-of-the-art (SOTA) methods on both DVD and GOPRO datasets and even yields more visually pleasing results in real video deblurring. Code and pre-trained models are publicly available at https://github.com/linjing7/VR-Baseline

Yuanhao Cai, Jing Lin, Xiaowan Hu et al.

Hyperspectral image (HSI) reconstruction aims to recover the 3D spatial-spectral signal from a 2D measurement in the coded aperture snapshot spectral imaging (CASSI) system. The HSI representations are highly similar and correlated across the spectral dimension. Modeling the inter-spectra interactions is beneficial for HSI reconstruction. However, existing CNN-based methods show limitations in capturing spectral-wise similarity and long-range dependencies. Besides, the HSI information is modulated by a coded aperture (physical mask) in CASSI. Nonetheless, current algorithms have not fully explored the guidance effect of the mask for HSI restoration. In this paper, we propose a novel framework, Mask-guided Spectral-wise Transformer (MST), for HSI reconstruction. Specifically, we present a Spectral-wise Multi-head Self-Attention (S-MSA) that treats each spectral feature as a token and calculates self-attention along the spectral dimension. In addition, we customize a Mask-guided Mechanism (MM) that directs S-MSA to pay attention to spatial regions with high-fidelity spectral representations. Extensive experiments show that our MST significantly outperforms state-of-the-art (SOTA) methods on simulation and real HSI datasets while requiring dramatically cheaper computational and memory costs. Code and pre-trained models are available at https://github.com/caiyuanhao1998/MST/

Ran Yu, Chenyu Tian, Weihao Xia et al.

Most existing video tasks related to "human" focus on the segmentation of salient humans, ignoring the unspecified others in the video. Few studies have focused on segmenting and tracking all humans in a complex video, including pedestrians and humans of other states (e.g., seated, riding, or occluded). In this paper, we propose a novel framework, abbreviated as HVISNet, that segments and tracks all presented people in given videos based on a one-stage detector. To better evaluate complex scenes, we offer a new benchmark called HVIS (Human Video Instance Segmentation), which comprises 1447 human instance masks in 805 high-resolution videos in diverse scenes. Extensive experiments show that our proposed HVISNet outperforms the state-of-the-art methods in terms of accuracy at a real-time inference speed (30 FPS), especially on complex video scenes. We also notice that using the center of the bounding box to distinguish different individuals severely deteriorates the segmentation accuracy, especially in heavily occluded conditions. This common phenomenon is referred to as the ambiguous positive samples problem. To alleviate this problem, we propose a mechanism named Inner Center Sampling to improve the accuracy of instance segmentation. Such a plug-and-play inner center sampling mechanism can be incorporated in any instance segmentation models based on a one-stage detector to improve the performance. In particular, it gains 4.1 mAP improvement on the state-of-the-art method in the case of occluded humans. Code and data are available at https://github.com/IIGROUP/HVISNet.

Chenyu Tian, Ran Yu, Xinyuan Zhao et al.

Current methods of multi-person pose estimation typically treat the localization and the association of body joints separately. It is convenient but inefficient, leading to additional computation and a waste of time. This paper, however, presents a novel framework PoseDet (Estimating Pose by Detection) to localize and associate body joints simultaneously at higher inference speed. Moreover, we propose the keypoint-aware pose embedding to represent an object in terms of the locations of its keypoints. The proposed pose embedding contains semantic and geometric information, allowing us to access discriminative and informative features efficiently. It is utilized for candidate classification and body joint localization in PoseDet, leading to robust predictions of various poses. This simple framework achieves an unprecedented speed and a competitive accuracy on the COCO benchmark compared with state-of-the-art methods. Extensive experiments on the CrowdPose benchmark show the robustness in the crowd scenes. Source code is available.

Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo et al.

In this paper, we propose a novel method called Residual Steps Network (RSN). RSN aggregates features with the same spatial size (Intra-level features) efficiently to obtain delicate local representations, which retain rich low-level spatial information and result in precise keypoint localization. Additionally, we observe the output features contribute differently to final performance. To tackle this problem, we propose an efficient attention mechanism - Pose Refine Machine (PRM) to make a trade-off between local and global representations in output features and further refine the keypoint locations. Our approach won the 1st place of COCO Keypoint Challenge 2019 and achieves state-of-the-art results on both COCO and MPII benchmarks, without using extra training data and pretrained model. Our single model achieves 78.6 on COCO test-dev, 93.0 on MPII test dataset. Ensembled models achieve 79.2 on COCO test-dev, 77.1 on COCO test-challenge dataset. The source code is publicly available for further research at https://github.com/caiyuanhao1998/RSN/