Angela Yao

Junbin Xiao, Pan Zhou, Angela Yao et al.

We propose to perform video question answering (VideoQA) in a Contrastive manner via a Video Graph Transformer model (CoVGT). CoVGT's uniqueness and superiority are three-fold: 1) It proposes a dynamic graph transformer module which encodes video by explicitly capturing the visual objects, their relations and dynamics, for complex spatio-temporal reasoning. 2) It designs separate video and text transformers for contrastive learning between the video and text to perform QA, instead of multi-modal transformer for answer classification. Fine-grained video-text communication is done by additional cross-modal interaction modules. 3) It is optimized by the joint fully- and self-supervised contrastive objectives between the correct and incorrect answers, as well as the relevant and irrelevant questions respectively. With superior video encoding and QA solution, we show that CoVGT can achieve much better performances than previous arts on video reasoning tasks. Its performances even surpass those models that are pretrained with millions of external data. We further show that CoVGT can also benefit from cross-modal pretraining, yet with orders of magnitude smaller data. The results demonstrate the effectiveness and superiority of CoVGT, and additionally reveal its potential for more data-efficient pretraining. We hope our success can advance VideoQA beyond coarse recognition/description towards fine-grained relation reasoning of video contents. Our code is available at https://github.com/doc-doc/CoVGT.

Guodong Ding, Fadime Sener, Angela Yao

Temporal action segmentation (TAS) in videos aims at densely identifying video frames in minutes-long videos with multiple action classes. As a long-range video understanding task, researchers have developed an extended collection of methods and examined their performance using various benchmarks. Despite the rapid growth of TAS techniques in recent years, no systematic survey has been conducted in these sectors. This survey analyzes and summarizes the most significant contributions and trends. In particular, we first examine the task definition, common benchmarks, types of supervision, and prevalent evaluation measures. In addition, we systematically investigate two essential techniques of this topic, i.e., frame representation and temporal modeling, which have been studied extensively in the literature. We then conduct a thorough review of existing TAS works categorized by their levels of supervision and conclude our survey by identifying and emphasizing several research gaps. In addition, we have curated a list of TAS resources, which is available at https://github.com/nus-cvml/awesome-temporal-action-segmentation.

Xuanlong Yu, Yi Zuo, Zitao Wang et al. · cmu

This paper outlines the winning solutions employed in addressing the MUAD uncertainty quantification challenge held at ICCV 2023. The challenge was centered around semantic segmentation in urban environments, with a particular focus on natural adversarial scenarios. The report presents the results of 19 submitted entries, with numerous techniques drawing inspiration from cutting-edge uncertainty quantification methodologies presented at prominent conferences in the fields of computer vision and machine learning and journals over the past few years. Within this document, the challenge is introduced, shedding light on its purpose and objectives, which primarily revolved around enhancing the robustness of semantic segmentation in urban scenes under varying natural adversarial conditions. The report then delves into the top-performing solutions. Moreover, the document aims to provide a comprehensive overview of the diverse solutions deployed by all participants. By doing so, it seeks to offer readers a deeper insight into the array of strategies that can be leveraged to effectively handle the inherent uncertainties associated with autonomous driving and semantic segmentation, especially within urban environments.

Bo Ji, Angela Yao

Video deblurring has achieved remarkable progress thanks to the success of deep neural networks. Most methods solve for the deblurring end-to-end with limited information propagation from the video sequence. However, different frame regions exhibit different characteristics and should be provided with corresponding relevant information. To achieve fine-grained deblurring, we designed a memory branch to memorize the blurry-sharp feature pairs in the memory bank, thus providing useful information for the blurry query input. To enrich the memory of our memory bank, we further designed a bidirectional recurrency and multi-scale strategy based on the memory bank. Experimental results demonstrate that our model outperforms other state-of-the-art methods while keeping the model complexity and inference time low. The code is available at https://github.com/jibo27/MemDeblur.

Yuehan Zhang, Bo Ji, Jia Hao et al.

In image super-resolution, both pixel-wise accuracy and perceptual fidelity are desirable. However, most deep learning methods only achieve high performance in one aspect due to the perception-distortion trade-off, and works that successfully balance the trade-off rely on fusing results from separately trained models with ad-hoc post-processing. In this paper, we propose a novel super-resolution model with a low-frequency constraint (LFc-SR), which balances the objective and perceptual quality through a single model and yields super-resolved images with high PSNR and perceptual scores. We further introduce an ADMM-based alternating optimization method for the non-trivial learning of the constrained model. Experiments showed that our method, without cumbersome post-processing procedures, achieved the state-of-the-art performance. The code is available at https://github.com/Yuehan717/PDASR.

Marwane Hariat, Olivier Laurent, Rémi Kazmierczak et al.

Semantic segmentation methods have advanced significantly. Still, their robustness to real-world perturbations and object types not seen during training remains a challenge, particularly in safety-critical applications. We propose a novel approach to improve the robustness of semantic segmentation techniques by leveraging the synergy between label-to-image generators and image-to-label segmentation models. Specifically, we design Robusta, a novel robust conditional generative adversarial network to generate realistic and plausible perturbed images that can be used to train reliable segmentation models. We conduct in-depth studies of the proposed generative model, assess the performance and robustness of the downstream segmentation network, and demonstrate that our approach can significantly enhance the robustness in the face of real-world perturbations, distribution shifts, and out-of-distribution samples. Our results suggest that this approach could be valuable in safety-critical applications, where the reliability of perception modules such as semantic segmentation is of utmost importance and comes with a limited computational budget in inference. We release our code at https://github.com/ENSTA-U2IS-AI/robusta.

Jiayin Zhu, Zhuoran Zhao, Linlin Yang et al.

We present HiFiHR, a high-fidelity hand reconstruction approach that utilizes render-and-compare in the learning-based framework from a single image, capable of generating visually plausible and accurate 3D hand meshes while recovering realistic textures. Our method achieves superior texture reconstruction by employing a parametric hand model with predefined texture assets, and by establishing a texture reconstruction consistency between the rendered and input images during training. Moreover, based on pretraining the network on an annotated dataset, we apply varying degrees of supervision using our pipeline, i.e., self-supervision, weak supervision, and full supervision, and discuss the various levels of contributions of the learned high-fidelity textures in enhancing hand pose and shape estimation. Experimental results on public benchmarks including FreiHAND and HO-3D demonstrate that our method outperforms the state-of-the-art hand reconstruction methods in texture reconstruction quality while maintaining comparable accuracy in pose and shape estimation. Our code is available at https://github.com/viridityzhu/HiFiHR.

Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov et al.

Assembly101 is a new procedural activity dataset featuring 4321 videos of people assembling and disassembling 101 "take-apart" toy vehicles. Participants work without fixed instructions, and the sequences feature rich and natural variations in action ordering, mistakes, and corrections. Assembly101 is the first multi-view action dataset, with simultaneous static (8) and egocentric (4) recordings. Sequences are annotated with more than 100K coarse and 1M fine-grained action segments, and 18M 3D hand poses. We benchmark on three action understanding tasks: recognition, anticipation and temporal segmentation. Additionally, we propose a novel task of detecting mistakes. The unique recording format and rich set of annotations allow us to investigate generalization to new toys, cross-view transfer, long-tailed distributions, and pose vs. appearance. We envision that Assembly101 will serve as a new challenge to investigate various activity understanding problems.

Hangyu Qin, Junbin Xiao, Angela Yao

This paper presents question-answering on dense video events, a novel task that answers and grounds dense-event questions in long videos, thus challenging MLLMs to faithfully comprehend and reason about multiple events over extended periods of time. To facilitate the study, we construct DeVE-QA -- a dataset featuring 78K questions about 26K events on 10.6K long videos. Our benchmarking shows that state-of-the-art MLLMs struggle on DeVE-QA. For improvement, we propose DeVi, a novel training-free MLLM approach that highlights a hierarchical captioning module, a temporal event memory module, and a self-consistency checking module to respectively detect, contextualize and memorize, and ground dense-events in long videos for question answering. Extensive experiments show that DeVi is superior at answering dense-event questions and grounding relevant video moments. Compared with existing MLLMs, it achieves a notable increase of 4.8% and 2.1% for G(round)QA accuracy on DeVE-QA and NExT-GQA, respectively. Data and code are available at https://github.com/QHUni/DeVE-QA.

Sheng Zhou, Junbin Xiao, Xun Yang et al.

Existing efforts in text-based video question answering (TextVideoQA) are criticized for their opaque decisionmaking and heavy reliance on scene-text recognition. In this paper, we propose to study Grounded TextVideoQA by forcing models to answer questions and spatio-temporally localize the relevant scene-text regions, thus decoupling QA from scenetext recognition and promoting research towards interpretable QA. The task has three-fold significance. First, it encourages scene-text evidence versus other short-cuts for answer predictions. Second, it directly accepts scene-text regions as visual answers, thus circumventing the problem of ineffective answer evaluation by stringent string matching. Third, it isolates the challenges inherited in VideoQA and scene-text recognition. This enables the diagnosis of the root causes for failure predictions, e.g., wrong QA or wrong scene-text recognition? To achieve Grounded TextVideoQA, we propose the T2S-QA model that highlights a disentangled temporal-to-spatial contrastive learning strategy for weakly-supervised scene-text grounding and grounded TextVideoQA. To facilitate evaluation, we construct a new dataset ViTXT-GQA which features 52K scene-text bounding boxes within 2.2K temporal segments related to 2K questions and 729 videos. With ViTXT-GQA, we perform extensive experiments and demonstrate the severe limitations of existing techniques in Grounded TextVideoQA. While T2S-QA achieves superior results, the large performance gap with human leaves ample space for improvement. Our further analysis of oracle scene-text inputs posits that the major challenge is scene-text recognition. To advance the research of Grounded TextVideoQA, our dataset and code are at https://github.com/zhousheng97/ViTXT-GQA.git

Dipika Singhania, Rahul Rahaman, Angela Yao

Temporal action segmentation tags action labels for every frame in an input untrimmed video containing multiple actions in a sequence. For the task of temporal action segmentation, we propose an encoder-decoder-style architecture named C2F-TCN featuring a "coarse-to-fine" ensemble of decoder outputs. The C2F-TCN framework is enhanced with a novel model agnostic temporal feature augmentation strategy formed by the computationally inexpensive strategy of the stochastic max-pooling of segments. It produces more accurate and well-calibrated supervised results on three benchmark action segmentation datasets. We show that the architecture is flexible for both supervised and representation learning. In line with this, we present a novel unsupervised way to learn frame-wise representation from C2F-TCN. Our unsupervised learning approach hinges on the clustering capabilities of the input features and the formation of multi-resolution features from the decoder's implicit structure. Further, we provide the first semi-supervised temporal action segmentation results by merging representation learning with conventional supervised learning. Our semi-supervised learning scheme, called ``Iterative-Contrastive-Classify (ICC)'', progressively improves in performance with more labeled data. The ICC semi-supervised learning in C2F-TCN, with 40% labeled videos, performs similar to fully supervised counterparts.

Kai Xu, Rongyu Chen, Gianni Franchi et al.

The capacity of a modern deep learning system to determine if a sample falls within its realm of knowledge is fundamental and important. In this paper, we offer insights and analyses of recent state-of-the-art out-of-distribution (OOD) detection methods - extremely simple activation shaping (ASH). We demonstrate that activation pruning has a detrimental effect on OOD detection, while activation scaling enhances it. Moreover, we propose SCALE, a simple yet effective post-hoc network enhancement method for OOD detection, which attains state-of-the-art OOD detection performance without compromising in-distribution (ID) accuracy. By integrating scaling concepts into the training process to capture a sample's ID characteristics, we propose Intermediate Tensor SHaping (ISH), a lightweight method for training time OOD detection enhancement. We achieve AUROC scores of +1.85\% for near-OOD and +0.74\% for far-OOD datasets on the OpenOOD v1.5 ImageNet-1K benchmark. Our code and models are available at https://github.com/kai422/SCALE.

Kai Xu, Ziwei Yu, Xin Wang et al.

In video super-resolution, it is common to use a frame-wise alignment to support the propagation of information over time. The role of alignment is well-studied for low-level enhancement in video, but existing works overlook a critical step -- resampling. We show through extensive experiments that for alignment to be effective, the resampling should preserve the reference frequency spectrum while minimizing spatial distortions. However, most existing works simply use a default choice of bilinear interpolation for resampling even though bilinear interpolation has a smoothing effect and hinders super-resolution. From these observations, we propose an implicit resampling-based alignment. The sampling positions are encoded by a sinusoidal positional encoding, while the value is estimated with a coordinate network and a window-based cross-attention. We show that bilinear interpolation inherently attenuates high-frequency information while an MLP-based coordinate network can approximate more frequencies. Experiments on synthetic and real-world datasets show that alignment with our proposed implicit resampling enhances the performance of state-of-the-art frameworks with minimal impact on both compute and parameters.

Haipeng Xiong, Angela Yao

Local counts, or the number of objects in a local area, is a continuous value by nature. Yet recent state-of-the-art methods show that formulating counting as a classification task performs better than regression. Through a series of experiments on carefully controlled synthetic data, we show that this counter-intuitive result is caused by imprecise ground truth local counts. Factors such as biased dot annotations and incorrectly matched Gaussian kernels used to generate ground truth counts introduce deviations from the true local counts. Standard continuous regression is highly sensitive to these errors, explaining the performance gap between classification and regression. To mitigate the sensitivity, we loosen the regression formulation from a continuous scale to a discrete ordering and propose a novel discrete-constrained (DC) regression. Applied to crowd counting, DC-regression is more accurate than both classification and standard regression on three public benchmarks. A similar advantage also holds for the age estimation task, verifying the overall effectiveness of DC-regression.

Kerui Gu, Linlin Yang, Michael Bi Mi et al.

In human and hand pose estimation, heatmaps are a crucial intermediate representation for a body or hand keypoint. Two popular methods to decode the heatmap into a final joint coordinate are via an argmax, as done in heatmap detection, or via softmax and expectation, as done in integral regression. Integral regression is learnable end-to-end, but has lower accuracy than detection. This paper uncovers an induced bias from integral regression that results from combining the softmax and the expectation operation. This bias often forces the network to learn degenerately localized heatmaps, obscuring the keypoint's true underlying distribution and leads to lower accuracies. Training-wise, by investigating the gradients of integral regression, we show that the implicit guidance of integral regression to update the heatmap makes it slower to converge than detection. To counter the above two limitations, we propose Bias Compensated Integral Regression (BCIR), an integral regression-based framework that compensates for the bias. BCIR also incorporates a Gaussian prior loss to speed up training and improve prediction accuracy. Experimental results on both the human body and hand benchmarks show that BCIR is faster to train and more accurate than the original integral regression, making it competitive with state-of-the-art detection methods.

Dibyadip Chatterjee, Fadime Sener, Shugao Ma et al.

Human actions in egocentric videos are often hand-object interactions composed from a verb (performed by the hand) applied to an object. Despite their extensive scaling up, egocentric datasets still face two limitations - sparsity of action compositions and a closed set of interacting objects. This paper proposes a novel open vocabulary action recognition task. Given a set of verbs and objects observed during training, the goal is to generalize the verbs to an open vocabulary of actions with seen and novel objects. To this end, we decouple the verb and object predictions via an object-agnostic verb encoder and a prompt-based object encoder. The prompting leverages CLIP representations to predict an open vocabulary of interacting objects. We create open vocabulary benchmarks on the EPIC-KITCHENS-100 and Assembly101 datasets; whereas closed-action methods fail to generalize, our proposed method is effective. In addition, our object encoder significantly outperforms existing open-vocabulary visual recognition methods in recognizing novel interacting objects.

Guodong Ding, Angela Yao

We present a semi-supervised learning approach to the temporal action segmentation task. The goal of the task is to temporally detect and segment actions in long, untrimmed procedural videos, where only a small set of videos are densely labelled, and a large collection of videos are unlabelled. To this end, we propose two novel loss functions for the unlabelled data: an action affinity loss and an action continuity loss. The action affinity loss guides the unlabelled samples learning by imposing the action priors induced from the labelled set. Action continuity loss enforces the temporal continuity of actions, which also provides frame-wise classification supervision. In addition, we propose an Adaptive Boundary Smoothing (ABS) approach to build coarser action boundaries for more robust and reliable learning. The proposed loss functions and ABS were evaluated on three benchmarks. Results show that they significantly improved action segmentation performance with a low amount (5% and 10%) of labelled data and achieved comparable results to full supervision with 50% labelled data. Furthermore, ABS succeeded in boosting performance when integrated into fully-supervised learning.

Guodong Ding, Fadime Sener, Shugao Ma et al.

One promising use case of AI assistants is to help with complex procedures like cooking, home repair, and assembly tasks. Can we teach the assistant to interject after the user makes a mistake? This paper targets the problem of identifying ordering mistakes in assembly procedures. We propose a system that can detect ordering mistakes by utilizing a learned knowledge base. Our framework constructs a knowledge base with spatial and temporal beliefs based on observed mistakes. Spatial beliefs depict the topological relationship of the assembling components, while temporal beliefs aggregate prerequisite actions as ordering constraints. With an episodic memory design, our algorithm can dynamically update and construct the belief sets as more actions are observed, all in an online fashion. We demonstrate experimentally that our inferred spatial and temporal beliefs are capable of identifying incorrect orderings in real-world action sequences. To construct the spatial beliefs, we collect a new set of coarse-level action annotations for Assembly101 based on the positioning of the toy parts. Finally, we demonstrate the superior performance of our belief inference algorithm in detecting ordering mistakes on the Assembly101 dataset.

Shaohui Lin, Bo Ji, Rongrong Ji et al.

Multi-exit architectures consist of a backbone and branch classifiers that offer shortened inference pathways to reduce the run-time of deep neural networks. In this paper, we analyze different branching patterns that vary in their allocation of computational complexity for the branch classifiers. Constant-complexity branching keeps all branches the same, while complexity-increasing and complexity-decreasing branching place more complex branches later or earlier in the backbone respectively. Through extensive experimentation on multiple backbones and datasets, we find that complexity-decreasing branches are more effective than constant-complexity or complexity-increasing branches, which achieve the best accuracy-cost trade-off. We investigate a cause by using knowledge consistency to probe the effect of adding branches onto a backbone. Our findings show that complexity-decreasing branching yields the least disruption to the feature abstraction hierarchy of the backbone, which explains the effectiveness of the branching patterns.

Xiaolu Liu, Yicong Li, Song Wang et al.

Recently, world models have been incorporated into the autonomous driving systems to improve the planning reliability. Existing approaches typically predict future states through appearance generation or deterministic regression, which limits their ability to capture trajectory-conditioned scene evolution and leads to unreliable action planning. To address this, we propose DynFlowDrive, a latent world model that leverages flow-based dynamics to model the transition of world states under different driving actions. By adopting the rectifiedflow formulation, the model learns a velocity field that describes how the scene state changes under different driving actions, enabling progressive prediction of future latent states. Building upon this, we further introduce a stability-aware multi-mode trajectory selection strategy that evaluates candidate trajectories according to the stability of the induced scene transitions. Extensive experiments on the nuScenes and NavSim benchmarks demonstrate consistent improvements across diverse driving frameworks without introducing additional inference overhead. Source code will be abaliable at https://github.com/xiaolul2/DynFlowDrive.

Ziwei Yu, Linlin Yang, You Xie et al.

We propose a novel framework for 3D hand shape reconstruction and hand-object grasp optimization from a single RGB image. The representation of hand-object contact regions is critical for accurate reconstructions. Instead of approximating the contact regions with sparse points, as in previous works, we propose a dense representation in the form of a UV coordinate map. Furthermore, we introduce inference-time optimization to fine-tune the grasp and improve interactions between the hand and the object. Our pipeline increases hand shape reconstruction accuracy and produces a vibrant hand texture. Experiments on datasets such as Ho3D, FreiHAND, and DexYCB reveal that our proposed method outperforms the state-of-the-art.

Yuehan Zhang, Angela Yao

In real-world video super-resolution (VSR), videos suffer from in-the-wild degradations and artifacts. VSR methods, especially recurrent ones, tend to propagate artifacts over time in the real-world setting and are more vulnerable than image super-resolution. This paper investigates the influence of artifacts on commonly used covariance-based attention mechanisms in VSR. Comparing the widely-used spatial attention, which computes covariance over space, versus the channel attention, we observe that the latter is less sensitive to artifacts. However, channel attention leads to feature redundancy, as evidenced by the higher covariance among output channels. As such, we explore simple techniques such as the squeeze-excite mechanism and covariance-based rescaling to counter the effects of high channel covariance. Based on our findings, we propose RealViformer. This channel-attention-based real-world VSR framework surpasses state-of-the-art on two real-world VSR datasets with fewer parameters and faster runtimes. The source code is available at https://github.com/Yuehan717/RealViformer.

Xinyao Liao, Qiyuan He, Kai Xu et al.

Autoregressive (AR) visual generation relies on tokenizers to map images to and from discrete sequences. However, tokenizers are trained to reconstruct clean images from ground-truth tokens, while AR generators are optimized only for token likelihood. This misalignment leads to generated token sequences that may decode into low-quality images, without direct supervision from the pixel space. We propose VA-$π$, a lightweight post-training framework that directly optimizes AR models with a principled pixel-space objective. VA-$π$ formulates the generator-tokenizer alignment as a variational optimization, deriving an evidence lower bound (ELBO) that unifies pixel reconstruction and autoregressive modeling. To optimize under the discrete token space, VA-$π$ introduces a reinforcement-based alignment strategy that treats the AR generator as a policy, uses pixel-space reconstruction quality as its intrinsic reward. The reward is measured by how well the predicted token sequences can reconstruct the original image under teacher forcing, giving the model direct pixel-level guidance without expensive free-running sampling. The regularization term of the ELBO serves as a natural regularizer, maintaining distributional consistency of tokens. VA-$π$ enables rapid adaptation of existing AR generators, without neither tokenizer retraining nor external reward models. With only 1% ImageNet-1K data and 25 minutes of tuning, it reduces FID from 14.36 to 7.65 and improves IS from 86.55 to 116.70 on LlamaGen-XXL, while also yielding notable gains in the text-to-image task on GenEval for both visual generation model (LlamaGen: from 0.306 to 0.339) and unified multi-modal model (Janus-Pro: from 0.725 to 0.744). Code is available at https://github.com/Lil-Shake/VA-Pi.

Xinyao Liao, Qiyuan He, Yicong Li et al.

Autoregressive image and video generators are trained with teacher-forced histories but must sample from their own generated prefixes at inference time, making them vulnerable to exposure bias and prefix drift. Existing remedies either modify training or apply sampling-time guidance aimed primarily at external semantic conditions, such as class labels or text prompts, rather than testing whether a next-step prediction provides strong posterior support for the generated prefix itself. We propose Visual Prefix Guidance (VPG), a training-free inference-time guidance method for autoregressive image and video generation. VPG improves next-step prediction by contrasting the model's output under the generated prefix with its output under a corrupted prefix, then extrapolating logits toward candidates that strengthen the posterior support of the generated prefix. Across class-conditional image generation with VAR, text-to-image generation with Infinity, and text-to-video generation with InfinityStar, VPG improves generation quality without retraining the base model, reducing FID on VAR by 0.36 on average and improving benchmark performance on both image and video generation.

You Xie, Huiqi Mao, Angela Yao et al.

We propose a novel approach to generate temporally coherent UV coordinates for loose clothing. Our method is not constrained by human body outlines and can capture loose garments and hair. We implemented a differentiable pipeline to learn UV mapping between a sequence of RGB inputs and textures via UV coordinates. Instead of treating the UV coordinates of each frame separately, our data generation approach connects all UV coordinates via feature matching for temporal stability. Subsequently, a generative model is trained to balance the spatial quality and temporal stability. It is driven by supervised and unsupervised losses in both UV and image spaces. Our experiments show that the trained models output high-quality UV coordinates and generalize to new poses. Once a sequence of UV coordinates has been inferred by our model, it can be used to flexibly synthesize new looks and modified visual styles. Compared to existing methods, our approach reduces the computational workload to animate new outfits by several orders of magnitude.

Shihao Zhang, Linlin Yang, Michael Bi Mi et al.

In computer vision, it is often observed that formulating regression problems as a classification task often yields better performance. We investigate this curious phenomenon and provide a derivation to show that classification, with the cross-entropy loss, outperforms regression with a mean squared error loss in its ability to learn high-entropy feature representations. Based on the analysis, we propose an ordinal entropy loss to encourage higher-entropy feature spaces while maintaining ordinal relationships to improve the performance of regression tasks. Experiments on synthetic and real-world regression tasks demonstrate the importance and benefits of increasing entropy for regression.

Junbin Xiao, Nanxin Huang, Hangyu Qin et al.

Video Large Language Models (Video-LLMs) are flourishing and has advanced many video-language tasks. As a golden testbed, Video Question Answering (VideoQA) plays pivotal role in Video-LLM developing. This work conducts a timely and comprehensive study of Video-LLMs' behavior in VideoQA, aiming to elucidate their success and failure modes, and provide insights towards more human-like video understanding and question answering. Our analyses demonstrate that Video-LLMs excel in VideoQA; they can correlate contextual cues and generate plausible responses to questions about varied video contents. However, models falter in handling video temporality, both in reasoning about temporal content ordering and grounding QA-relevant temporal moments. Moreover, the models behave unintuitively - they are unresponsive to adversarial video perturbations while being sensitive to simple variations of candidate answers and questions. Also, they do not necessarily generalize better. The findings demonstrate Video-LLMs' QA capability in standard condition yet highlight their severe deficiency in robustness and interpretability, suggesting the urgent need on rationales in Video-LLM developing.

Rahul Rahaman, Dipika Singhania, Alexandre Thiery et al.

In temporal action segmentation, Timestamp supervision requires only a handful of labelled frames per video sequence. For unlabelled frames, previous works rely on assigning hard labels, and performance rapidly collapses under subtle violations of the annotation assumptions. We propose a novel Expectation-Maximization (EM) based approach that leverages the label uncertainty of unlabelled frames and is robust enough to accommodate possible annotation errors. With accurate timestamp annotations, our proposed method produces SOTA results and even exceeds the fully-supervised setup in several metrics and datasets. When applied to timestamp annotations with missing action segments, our method presents stable performance. To further test our formulation's robustness, we introduce the new challenging annotation setup of Skip-tag supervision. This setup relaxes constraints and requires annotations of any fixed number of random frames in a video, making it more flexible than Timestamp supervision while remaining competitive.

Yuehan Zhang, Seungjun Lee, Angela Yao

Standard single-image super-resolution creates paired training data from high-resolution images through fixed downsampling kernels. However, real-world super-resolution (RWSR) faces unknown degradations in the low-resolution inputs, all the while lacking paired training data. Existing methods approach this problem by learning blind general models through complex synthetic augmentations on training inputs; they sacrifice the performance on specific degradation for broader generalization to many possible ones. We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel pairwise distance distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and inter-model distances across the specialized model and an auxiliary generalized model. Experiments on diverse datasets demonstrate that our method significantly enhances fidelity and perceptual quality, surpassing state-of-the-art approaches in RWSR. The source code is available at https://github.com/Yuehan717/PDD.

Xiaolu Liu, Yicong Li, Qiyuan He et al.

Textured 3D morphing seeks to generate smooth and plausible transitions between two 3D assets, preserving both structural coherence and fine-grained appearance. This ability is crucial not only for advancing 3D generation research but also for practical applications in animation, editing, and digital content creation. Existing approaches either operate directly on geometry, limiting them to shape-only morphing while neglecting textures, or extend 2D interpolation strategies into 3D, which often causes semantic ambiguity, structural misalignment, and texture blurring. These challenges underscore the necessity to jointly preserve geometric consistency, texture alignment, and robustness throughout the transition process. To address this, we propose Interp3D, a novel training-free framework for textured 3D morphing. It harnesses generative priors and adopts a progressive alignment principle to ensure both geometric fidelity and texture coherence. Starting from semantically aligned interpolation in condition space, Interp3D enforces structural consistency via SLAT (Structured Latent)-guided structure interpolation, and finally transfers appearance details through fine-grained texture fusion. For comprehensive evaluations, we construct a dedicated dataset, Interp3DData, with graded difficulty levels and assess generation results from fidelity, transition smoothness, and plausibility. Both quantitative metrics and human studies demonstrate the significant advantages of our proposed approach over previous methods. Source code is available at https://github.com/xiaolul2/Interp3D.

Zhanzhong Pang, Fadime Sener, Shrinivas Ramasubramanian et al.

Procedural activity videos often exhibit a long-tailed action distribution due to varying action frequencies and durations. However, state-of-the-art temporal action segmentation methods overlook the long tail and fail to recognize tail actions. Existing long-tail methods make class-independent assumptions and struggle to identify tail classes when applied to temporal segmentation frameworks. This work proposes a novel group-wise temporal logit adjustment~(G-TLA) framework that combines a group-wise softmax formulation while leveraging activity information and action ordering for logit adjustment. The proposed framework significantly improves in segmenting tail actions without any performance loss on head actions.

Tanuj Sur, Shashank Tripathi, Nikos Athanasiou et al.

We introduce UniCon3R (Unified Contact-aware 3D Reconstruction), a unified feed-forward framework for online human-scene 4D reconstruction from monocular videos. Recent feed-forward methods enable real-time world-coordinate human motion and scene reconstruction, but they often produce physically implausible artifacts such as bodies floating above the ground or penetrating parts of the scene. The key reason is that existing approaches fail to model physical interactions between the human and the environment. A natural next step is to predict human-scene contact as an auxiliary output -- yet we find this alone is not sufficient: contact must actively correct the reconstruction. To address this, we explicitly model interaction by inferring 3D contact from the human pose and scene geometry and use the contact as a corrective cue for generating the final pose. This enables UniCon3R to jointly recover high-fidelity scene geometry and spatially aligned 3D humans within the scene. Experiments on standard human-centric video benchmarks such as RICH, EMDB, 3DPW and SLOPER4D show that UniCon3R outperforms state-of-the-art baselines on physical plausibility and global human motion estimation while achieving real-time online inference. We experimentally demonstrate that contact serves as a powerful internal prior rather than just an external metric, thus establishing a new paradigm for physically grounded joint human-scene reconstruction. Project page is available at https://surtantheta.github.io/UniCon3R .

Zhongqun Zhang, Hengfei Wang, Ziwei Yu et al.

Modeling the physical contacts between the hand and object is standard for refining inaccurate hand poses and generating novel human grasp in 3D hand-object reconstruction. However, existing methods rely on geometric constraints that cannot be specified or controlled. This paper introduces a novel task of controllable 3D hand-object contact modeling with natural language descriptions. Challenges include i) the complexity of cross-modal modeling from language to contact, and ii) a lack of descriptive text for contact patterns. To address these issues, we propose NL2Contact, a model that generates controllable contacts by leveraging staged diffusion models. Given a language description of the hand and contact, NL2Contact generates realistic and faithful 3D hand-object contacts. To train the model, we build \textit{ContactDescribe}, the first dataset with hand-centered contact descriptions. It contains multi-level and diverse descriptions generated by large language models based on carefully designed prompts (e.g., grasp action, grasp type, contact location, free finger status). We show applications of our model to grasp pose optimization and novel human grasp generation, both based on a textual contact description.

Haipeng Xiong, Angela Yao

Regression tasks in computer vision, such as age estimation or counting, are often formulated into classification by quantizing the target space into classes. Yet real-world data is often imbalanced -- the majority of training samples lie in a head range of target values, while a minority of samples span a usually larger tail range. By selecting the class quantization, one can adjust imbalanced regression targets into balanced classification outputs, though there are trade-offs in balancing classification accuracy and quantization error. To improve regression performance over the entire range of data, we propose to construct hierarchical classifiers for solving imbalanced regression tasks. The fine-grained classifiers limit the quantization error while being modulated by the coarse predictions to ensure high accuracy. Standard hierarchical classification approaches, however, when applied to the regression problem, fail to ensure that predicted ranges remain consistent across the hierarchy. As such, we propose a range-preserving distillation process that can effectively learn a single classifier from the set of hierarchical classifiers. Our novel hierarchical classification adjustment (HCA) for imbalanced regression shows superior results on three diverse tasks: age estimation, crowd counting and depth estimation. We will release the source code upon acceptance.

Junbin Xiao, Jiajun Chen, Tianxiang Sun et al.

Long streaming video QA remains challenging due to growing visual tokens and limited reasoning length of large language models (LLMs). KV-caching stores the Key-Value (KV) of the historical tokens via LLM prefill and enables more efficient streaming QA. However, existing methods cache every one or two frames, causing redundant memory usage and losing fine-grained spatial details within frame or temporal contexts across frames. This paper proposes MuKV, a method that features a multi-grained KV cache compression module and a semi-hierarchical retrieval approach to improve both efficiency and accuracy for long streaming VideoQA. For the offline KV cache, MuKV extracts visual representations at patch-, frame-, and segment-levels. The multiple levels of granularity preserve both local cues and global temporal context, while maintaining efficiency with a dual signal token compression mechanism guided by self-attention and frequency. For online QA, MuKV designs a semi-hierarchical retrieval method to retrieve relevant KV caches for answer generation. Experiments on long-streaming VideoQA benchmarks show that MuKV significantly improves answer accuracy, without sacrificing memory and online QA efficiency. Moreover, our compression mechanism alone brings consistent benefits across answer accuracy, memory, and QA efficiency over baselines, showcasing highly effective contribution.

Sheng Zhou, Junbin Xiao, Qingyun Li et al.

We introduce EgoTextVQA, a novel and rigorously constructed benchmark for egocentric QA assistance involving scene text. EgoTextVQA contains 1.5K ego-view videos and 7K scene-text aware questions that reflect real user needs in outdoor driving and indoor house-keeping activities. The questions are designed to elicit identification and reasoning on scene text in an egocentric and dynamic environment. With EgoTextVQA, we comprehensively evaluate 10 prominent multimodal large language models. Currently, all models struggle, and the best results (Gemini 1.5 Pro) are around 33\% accuracy, highlighting the severe deficiency of these techniques in egocentric QA assistance. Our further investigations suggest that precise temporal grounding and multi-frame reasoning, along with high resolution and auxiliary scene-text inputs, are key for better performance. With thorough analyses and heuristic suggestions, we hope EgoTextVQA can serve as a solid testbed for research in egocentric scene-text QA assistance. Our dataset is released at: https://github.com/zhousheng97/EgoTextVQA.

Qiyuan He, Jinghao Wang, Ziwei Liu et al.

Conditional diffusion models can create unseen images in various settings, aiding image interpolation. Interpolation in latent spaces is well-studied, but interpolation with specific conditions like text or poses is less understood. Simple approaches, such as linear interpolation in the space of conditions, often result in images that lack consistency, smoothness, and fidelity. To that end, we introduce a novel training-free technique named Attention Interpolation via Diffusion (AID). Our key contributions include 1) proposing an inner/outer interpolated attention layer; 2) fusing the interpolated attention with self-attention to boost fidelity; and 3) applying beta distribution to selection to increase smoothness. We also present a variant, Prompt-guided Attention Interpolation via Diffusion (PAID), that considers interpolation as a condition-dependent generative process. This method enables the creation of new images with greater consistency, smoothness, and efficiency, and offers control over the exact path of interpolation. Our approach demonstrates effectiveness for conceptual and spatial interpolation. Code and demo are available at https://github.com/QY-H00/attention-interpolation-diffusion.

Zhanzhong Pang, Dibyadip Chatterjee, Fadime Sener et al.

Multimodal Large Language Models (MLLMs) have advanced open-world action understanding and can be adapted as generative classifiers for closed-set settings by autoregressively generating action labels as text. However, this approach is inefficient, and shared subwords across action labels introduce semantic overlap, leading to ambiguity in generation. In contrast, discriminative classifiers learn task-specific representations with clear decision boundaries, enabling efficient one-step classification without autoregressive decoding. We first compare generative and discriminative classifiers with MLLMs for closed-set action understanding, revealing the superior accuracy and efficiency of the latter. To bridge the performance gap, we design strategies that elevate generative classifiers toward performance comparable with discriminative ones. Furthermore, we show that generative modeling can complement discriminative classifiers, leading to better performance while preserving efficiency. To this end, we propose Generation-Assisted Discriminative~(GAD) classifier for closed-set action understanding. GAD operates only during fine-tuning, preserving full compatibility with MLLM pretraining. Extensive experiments on temporal action understanding benchmarks demonstrate that GAD improves both accuracy and efficiency over generative methods, achieving state-of-the-art results on four tasks across five datasets, including an average 2.5% accuracy gain and 3x faster inference on our largest COIN benchmark.

Haotian Ye, Qiyuan He, Jiaqi Han et al.

Accurate and efficient discrete video tokenization is essential for long video sequences processing. Yet, the inherent complexity and variable information density of videos present a significant bottleneck for current tokenizers, which rigidly compress all content at a fixed rate, leading to redundancy or information loss. Drawing inspiration from Shannon's information theory, this paper introduces InfoTok, a principled framework for adaptive video tokenization. We rigorously prove that existing data-agnostic training methods are suboptimal in representation length, and present a novel evidence lower bound (ELBO)-based algorithm that approaches theoretical optimality. Leveraging this framework, we develop a transformer-based adaptive compressor that enables adaptive tokenization. Empirical results demonstrate state-of-the-art compression performance, saving 20% tokens without influence on performance, and achieving 2.3x compression rates while still outperforming prior heuristic adaptive approaches. By allocating tokens according to informational richness, InfoTok enables a more compressed yet accurate tokenization for video representation, offering valuable insights for future research.

Kerui Gu, Rongyu Chen, Angela Yao

Most 2D human pose estimation frameworks estimate keypoint confidence in an ad-hoc manner, using heuristics such as the maximum value of heatmaps. The confidence is part of the evaluation scheme, e.g., AP for the MSCOCO dataset, yet has been largely overlooked in the development of state-of-the-art methods. This paper takes the first steps in addressing miscalibration in pose estimation. From a calibration point of view, the confidence should be aligned with the pose accuracy. In practice, existing methods are poorly calibrated. We show, through theoretical analysis, why a miscalibration gap exists and how to narrow the gap. Simply predicting the instance size and adjusting the confidence function gives considerable AP improvements. Given the black-box nature of deep neural networks, however, it is not possible to fully close this gap with only closed-form adjustments. As such, we go one step further and learn network-specific adjustments by enforcing consistency between confidence and pose accuracy. Our proposed Calibrated ConfidenceNet (CCNet) is a light-weight post-hoc addition that improves AP by up to 1.4% on off-the-shelf pose estimation frameworks. Applied to the downstream task of mesh recovery, CCNet facilitates an additional 1.0mm decrease in 3D keypoint error.

Zhuo Chen, Fanyue Wei, Runze Xu et al.

Training-free image editing with large diffusion models has become practical, yet faithfully performing complex non-rigid edits (e.g., pose or shape changes) remains highly challenging. We identify a key underlying cause: attention collapse in existing attention sharing mechanisms, where either positional embeddings or semantic features dominate visual content retrieval, leading to over-editing or under-editing. To address this issue, we introduce SynPS, a method that Synergistically leverages Positional embeddings and Semantic information for faithful non-rigid image editing. We first propose an editing measurement that quantifies the required editing magnitude at each denoising step. Based on this measurement, we design an attention synergy pipeline that dynamically modulates the influence of positional embeddings, enabling SynPS to balance semantic modifications and fidelity preservation. By adaptively integrating positional and semantic cues, SynPS effectively avoids both over- and under-editing. Extensive experiments on public and newly curated benchmarks demonstrate the superior performance and faithfulness of our approach.

Joseph Hoche, Andrei Bursuc, David Brellmann et al.

Large Vision-Language Models (LVLMs) often produce plausible but unreliable outputs, making robust uncertainty estimation essential. Recent work on semantic uncertainty estimates relies on external models to cluster multiple sampled responses and measure their semantic consistency. However, these clustering methods are often fragile, highly sensitive to minor phrasing variations, and can incorrectly group or separate semantically similar answers, leading to unreliable uncertainty estimates. We propose Semantic Gaussian Process Uncertainty (SGPU), a Bayesian framework that quantifies semantic uncertainty by analyzing the geometric structure of answer embeddings, avoiding brittle clustering. SGPU maps generated answers into a dense semantic space, computes the Gram matrix of their embeddings, and summarizes their semantic configuration via the eigenspectrum. This spectral representation is then fed into a Gaussian Process Classifier that learns to map patterns of semantic consistency to predictive uncertainty, and that can be applied in both black-box and white-box settings. Across six LLMs and LVLMs on eight datasets spanning VQA, image classification, and textual QA, SGPU consistently achieves state-of-the-art calibration (ECE) and discriminative (AUROC, AUARC) performance. We further show that SGPU transfers across models and modalities, indicating that its spectral representation captures general patterns of semantic uncertainty.

Bo Ji, Angela Yao

Standard 3D Gaussian Splatting (3DGS) relies on known or pre-computed camera poses and a sparse point cloud, obtained from structure-from-motion (SfM) preprocessing, to initialize and grow 3D Gaussians. We propose a novel SfM-Free 3DGS (SFGS) method for video input, eliminating the need for known camera poses and SfM preprocessing. Our approach introduces a hierarchical training strategy that trains and merges multiple 3D Gaussian representations -- each optimized for specific scene regions -- into a single, unified 3DGS model representing the entire scene. To compensate for large camera motions, we leverage video frame interpolation models. Additionally, we incorporate multi-source supervision to reduce overfitting and enhance representation. Experimental results reveal that our approach significantly surpasses state-of-the-art SfM-free novel view synthesis methods. On the Tanks and Temples dataset, we improve PSNR by an average of 2.25dB, with a maximum gain of 3.72dB in the best scene. On the CO3D-V2 dataset, we achieve an average PSNR boost of 1.74dB, with a top gain of 3.90dB. The code is available at https://github.com/jibo27/3DGS_Hierarchical_Training.

Junbin Xiao, Nanxin Huang, Hao Qiu et al.

We present EgoBlind, the first egocentric VideoQA dataset collected from blind individuals to evaluate the assistive capabilities of contemporary multimodal large language models (MLLMs). EgoBlind comprises 1,392 first-person videos from the daily lives of blind and visually impaired individuals. It also features 5,311 questions directly posed or verified by the blind to reflect their in-situation needs for visual assistance. Each question has an average of 3 manually annotated reference answers to reduce subjectiveness. Using EgoBlind, we comprehensively evaluate 16 advanced MLLMs and find that all models struggle. The best performers achieve an accuracy near 60\%, which is far behind human performance of 87.4\%. To guide future advancements, we identify and summarize major limitations of existing MLLMs in egocentric visual assistance for the blind and explore heuristic solutions for improvement. With these efforts, we hope that EgoBlind will serve as a foundation for developing effective AI assistants to enhance the independence of the blind and visually impaired. Data and code are available at https://github.com/doc-doc/EgoBlind.

Minjoon Jung, Junbin Xiao, Byoung-Tak Zhang et al.

Video large language models (Video-LLMs) can temporally ground language queries and retrieve video moments. Yet, such temporal comprehension capabilities are neither well-studied nor understood. So we conduct a study on prediction consistency -- a key indicator for robustness and trustworthiness of temporal grounding. After the model identifies an initial moment within the video content, we apply a series of probes to check if the model's responses align with this initial grounding as an indicator of reliable comprehension. Our results reveal that current Video-LLMs are sensitive to variations in video contents, language queries, and task settings, unveiling severe deficiencies in maintaining consistency. We further explore common prompting and instruction-tuning methods as potential solutions, but find that their improvements are often unstable. To that end, we propose event temporal verification tuning that explicitly accounts for consistency, and demonstrate significant improvements for both grounding and consistency. Our data and code are open-sourced at https://github.com/minjoong507/Consistency-of-Video-LLM.

Qiyuan He, Angela Yao

Personalized image generation with text-to-image diffusion models generates unseen images based on reference image content. Zero-shot adapter methods such as IP-Adapter and OminiControl are especially interesting because they do not require test-time fine-tuning. However, they struggle to balance preserving personalized content and adherence to the text prompt. We identify a critical design flaw resulting in this performance gap: current adapters inadequately integrate personalization images with the textual descriptions. The generated images, therefore, replicate the personalized content rather than adhere to the text prompt instructions. Yet the base text-to-image has strong conceptual understanding capabilities that can be leveraged. We propose Conceptrol, a simple yet effective framework that enhances zero-shot adapters without adding computational overhead. Conceptrol constrains the attention of visual specification with a textual concept mask that improves subject-driven generation capabilities. It achieves as much as 89% improvement on personalization benchmarks over the vanilla IP-Adapter and can even outperform fine-tuning approaches such as Dreambooth LoRA. The source code is available at https://github.com/QY-H00/Conceptrol.

Junbin Xiao, Shenglang Zhang, Pengxiang Zhu et al.

We present the first systematic analysis of multimodal large language models (MLLMs) in personalized question-answering requiring ego-grounding - the ability to understand the camera-wearer in egocentric videos. To this end, we introduce MyEgo, the first egocentric VideoQA dataset designed to evaluate MLLMs' ability to understand, remember, and reason about the camera wearer. MyEgo comprises 541 long videos and 5K personalized questions asking about "my things", "my activities", and "my past". Benchmarking reveals that competitive MLLMs across variants, including open-source vs. proprietary, thinking vs. non-thinking, small vs. large scales all struggle on MyEgo. Top closed- and open-source models (e.g., GPT-5 and Qwen3-VL) achieve only~46% and 36% accuracy, trailing human performance by near 40% and 50% respectively. Surprisingly, neither explicit reasoning nor model scaling yield consistent improvements. Models improve when relevant evidence is explicitly provided, but gains drop over time, indicating limitations in tracking and remembering "me" and "my past". These findings collectively highlight the crucial role of ego-grounding and long-range memory in enabling personalized QA in egocentric videos. We hope MyEgo and our analyses catalyze further progress in these areas for egocentric personalized assistance. Data and code are available at https://github.com/Ryougetsu3606/MyEgo

Shihao Zhang, kenji kawaguchi, Angela Yao

Most works studying representation learning focus only on classification and neglect regression. Yet, the learning objectives and, therefore, the representation topologies of the two tasks are fundamentally different: classification targets class separation, leading to disconnected representations, whereas regression requires ordinality with respect to the target, leading to continuous representations. We thus wonder how the effectiveness of a regression representation is influenced by its topology, with evaluation based on the Information Bottleneck (IB) principle. The IB principle is an important framework that provides principles for learning effective representations. We establish two connections between it and the topology of regression representations. The first connection reveals that a lower intrinsic dimension of the feature space implies a reduced complexity of the representation Z. This complexity can be quantified as the conditional entropy of Z on the target Y, and serves as an upper bound on the generalization error. The second connection suggests a feature space that is topologically similar to the target space will better align with the IB principle. Based on these two connections, we introduce PH-Reg, a regularizer specific to regression that matches the intrinsic dimension and topology of the feature space with the target space. Experiments on synthetic and real-world regression tasks demonstrate the benefits of PH-Reg. Code: https://github.com/needylove/PH-Reg.

Qinchuan Cheng, Zhantao Gong, Pengzhan Sun et al.

Embodied agents in household environments must plan under partial observation: they need to remember objects, track state changes, and recover when actions fail. Existing benchmarks only partially test this ability. Egocentric video datasets capture realistic human activities but remain passive, while interactive simulators support execution but rely on synthetic scenes and hand-crafted dynamics, introducing a sim-to-real gap and often assuming fully observable state. We introduce Ego2World, an executable benchmark that turns egocentric cooking videos into executable symbolic worlds governed by graph-transition rules. Built on HD-EPIC, Ego2World derives reusable transition rules from video annotations and executes them in a hidden symbolic world graph. During evaluation, the simulator maintains the hidden world graph, while the agent plans over its own partial belief graph using only local observations and execution feedback. This separation forces agents to update memory and replan without observing the true world state. Experiments show that action-overlap scores overestimate physical-state success, and that persistent belief memory improves task completion while reducing repeated visual exploration -- suggesting that belief maintenance should be a first-class target of embodied-agent evaluation.

Huixin Sun, Linlin Yang, Ronyu Chen et al.

Despite significant advances in generic object detection, a persistent performance gap remains for tiny objects compared to normal-scale objects. We demonstrate that tiny objects are highly sensitive to annotation noise, where optimizing strict localization objectives risks noise overfitting. To address this, we propose Tiny Object Localization with Flows (TOLF), a noise-robust localization framework leveraging normalizing flows for flexible error modeling and uncertainty-guided optimization. Our method captures complex, non-Gaussian prediction distributions through flow-based error modeling, enabling robust learning under noisy supervision. An uncertainty-aware gradient modulation mechanism further suppresses learning from high-uncertainty, noise-prone samples, mitigating overfitting while stabilizing training. Extensive experiments across three datasets validate our approach's effectiveness. Especially, TOLF boosts the DINO baseline by 1.2% AP on the AI-TOD dataset.