Hehe Fan

Tao Zhuo, Zhiyong Cheng, Hehe Fan et al.

Continual Learning (CL) aims to incrementally update a trained model on new tasks without forgetting the acquired knowledge of old ones. Existing CL methods usually reduce forgetting with task priors, \ie using task identity or a subset of previously seen samples for model training. However, these methods would be infeasible when such priors are unknown in real-world applications. To address this fundamental but seldom-studied problem, we propose a Prior-Free Continual Learning (PFCL) method, which learns new tasks without knowing the task identity or any previous data. First, based on a fixed single-head architecture, we eliminate the need for task identity to select the task-specific output head. Second, we employ a regularization-based strategy for consistent predictions between the new and old models, avoiding revisiting previous samples. However, using this strategy alone often performs poorly in class-incremental scenarios, particularly for a long sequence of tasks. By analyzing the effectiveness and limitations of conventional regularization-based methods, we propose enhancing model consistency with an auxiliary unlabeled dataset additionally. Moreover, since some auxiliary data may degrade the performance, we further develop a reliable sample selection strategy to obtain consistent performance improvement. Extensive experiments on multiple image classification benchmark datasets show that our PFCL method significantly mitigates forgetting in all three learning scenarios. Furthermore, when compared to the most recent rehearsal-based methods that replay a limited number of previous samples, PFCL achieves competitive accuracy. Our code is available at: https://github.com/visiontao/pfcl

Hehe Fan, Xin Yu, Yuhang Ding et al.

Point cloud sequences are irregular and unordered in the spatial dimension while exhibiting regularities and order in the temporal dimension. Therefore, existing grid based convolutions for conventional video processing cannot be directly applied to spatio-temporal modeling of raw point cloud sequences. In this paper, we propose a point spatio-temporal (PST) convolution to achieve informative representations of point cloud sequences. The proposed PST convolution first disentangles space and time in point cloud sequences. Then, a spatial convolution is employed to capture the local structure of points in the 3D space, and a temporal convolution is used to model the dynamics of the spatial regions along the time dimension. Furthermore, we incorporate the proposed PST convolution into a deep network, namely PSTNet, to extract features of point cloud sequences in a hierarchical manner. Extensive experiments on widely-used 3D action recognition and 4D semantic segmentation datasets demonstrate the effectiveness of PSTNet to model point cloud sequences.

Zhiqiang Shen, Xiaoxiao Sheng, Hehe Fan et al.

Recently, the community has made tremendous progress in developing effective methods for point cloud video understanding that learn from massive amounts of labeled data. However, annotating point cloud videos is usually notoriously expensive. Moreover, training via one or only a few traditional tasks (e.g., classification) may be insufficient to learn subtle details of the spatio-temporal structure existing in point cloud videos. In this paper, we propose a Masked Spatio-Temporal Structure Prediction (MaST-Pre) method to capture the structure of point cloud videos without human annotations. MaST-Pre is based on spatio-temporal point-tube masking and consists of two self-supervised learning tasks. First, by reconstructing masked point tubes, our method is able to capture the appearance information of point cloud videos. Second, to learn motion, we propose a temporal cardinality difference prediction task that estimates the change in the number of points within a point tube. In this way, MaST-Pre is forced to model the spatial and temporal structure in point cloud videos. Extensive experiments on MSRAction-3D, NTU-RGBD, NvGesture, and SHREC'17 demonstrate the effectiveness of the proposed method.

Guangzhi Wang, Hehe Fan, Mohan Kankanhalli

Automatically localizing a position based on a few natural language instructions is essential for future robots to communicate and collaborate with humans. To approach this goal, we focus on the text-to-point-cloud cross-modal localization problem. Given a textual query, it aims to identify the described location from city-scale point clouds. The task involves two challenges. 1) In city-scale point clouds, similar ambient instances may exist in several locations. Searching each location in a huge point cloud with only instances as guidance may lead to less discriminative signals and incorrect results. 2) In textual descriptions, the hints are provided separately. In this case, the relations among those hints are not explicitly described, leading to difficulties of learning relations. To overcome these two challenges, we propose a unified Relation-Enhanced Transformer (RET) to improve representation discriminability for both point cloud and natural language queries. The core of the proposed RET is a novel Relation-enhanced Self-Attention (RSA) mechanism, which explicitly encodes instance (hint)-wise relations for the two modalities. Moreover, we propose a fine-grained cross-modal matching method to further refine the location predictions in a subsequent instance-hint matching stage. Experimental results on the KITTI360Pose dataset demonstrate that our approach surpasses the previous state-of-the-art method by large margin.

Wenjin Hou, Dingjie Fu, Kun Li et al.

Zero-shot learning (ZSL) aims to recognize unseen classes by transferring semantic knowledge from seen classes to unseen ones, guided by semantic information. To this end, existing works have demonstrated remarkable performance by utilizing global visual features from Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs) for visual-semantic interactions. Due to the limited receptive fields of CNNs and the quadratic complexity of ViTs, however, these visual backbones achieve suboptimal visual-semantic interactions. In this paper, motivated by the visual state space model (i.e., Vision Mamba), which is capable of capturing long-range dependencies and modeling complex visual dynamics, we propose a parameter-efficient ZSL framework called ZeroMamba to advance ZSL. Our ZeroMamba comprises three key components: Semantic-aware Local Projection (SLP), Global Representation Learning (GRL), and Semantic Fusion (SeF). Specifically, SLP integrates semantic embeddings to map visual features to local semantic-related representations, while GRL encourages the model to learn global semantic representations. SeF combines these two semantic representations to enhance the discriminability of semantic features. We incorporate these designs into Vision Mamba, forming an end-to-end ZSL framework. As a result, the learned semantic representations are better suited for classification. Through extensive experiments on four prominent ZSL benchmarks, ZeroMamba demonstrates superior performance, significantly outperforming the state-of-the-art (i.e., CNN-based and ViT-based) methods under both conventional ZSL (CZSL) and generalized ZSL (GZSL) settings. Code is available at: https://anonymous.4open.science/r/ZeroMamba.

Yi Wang, Zhiwen Fan, Tianlong Chen et al.

Vision Transformers (ViTs) have proven to be effective, in solving 2D image understanding tasks by training over large-scale image datasets; and meanwhile as a somehow separate track, in modeling the 3D visual world too such as voxels or point clouds. However, with the growing hope that transformers can become the "universal" modeling tool for heterogeneous data, ViTs for 2D and 3D tasks have so far adopted vastly different architecture designs that are hardly transferable. That invites an (over-)ambitious question: can we close the gap between the 2D and 3D ViT architectures? As a piloting study, this paper demonstrates the appealing promise to understand the 3D visual world, using a standard 2D ViT architecture, with only minimal customization at the input and output levels without redesigning the pipeline. To build a 3D ViT from its 2D sibling, we "inflate" the patch embedding and token sequence, accompanied with new positional encoding mechanisms designed to match the 3D data geometry. The resultant "minimalist" 3D ViT, named Simple3D-Former, performs surprisingly robustly on popular 3D tasks such as object classification, point cloud segmentation and indoor scene detection, compared to highly customized 3D-specific designs. It can hence act as a strong baseline for new 3D ViTs. Moreover, we note that pursing a unified 2D-3D ViT design has practical relevance besides just scientific curiosity. Specifically, we demonstrate that Simple3D-Former naturally enables to exploit the wealth of pre-trained weights from large-scale realistic 2D images (e.g., ImageNet), which can be plugged in to enhancing the 3D task performance "for free".

Xiaoyu Feng, Heming Du, Yueqi Duan et al.

Effectively preserving and encoding structure features from objects in irregular and sparse LiDAR points is a key challenge to 3D object detection on point cloud. Recently, Transformer has demonstrated promising performance on many 2D and even 3D vision tasks. Compared with the fixed and rigid convolution kernels, the self-attention mechanism in Transformer can adaptively exclude the unrelated or noisy points and thus suitable for preserving the local spatial structure in irregular LiDAR point cloud. However, Transformer only performs a simple sum on the point features, based on the self-attention mechanism, and all the points share the same transformation for value. Such isotropic operation lacks the ability to capture the direction-distance-oriented local structure which is important for 3D object detection. In this work, we propose a Structure-Embedding transFormer (SEFormer), which can not only preserve local structure as traditional Transformer but also have the ability to encode the local structure. Compared to the self-attention mechanism in traditional Transformer, SEFormer learns different feature transformations for value points based on the relative directions and distances to the query point. Then we propose a SEFormer based network for high-performance 3D object detection. Extensive experiments show that the proposed architecture can achieve SOTA results on Waymo Open Dataset, the largest 3D detection benchmark for autonomous driving. Specifically, SEFormer achieves 79.02% mAP, which is 1.2% higher than existing works. We will release the codes.

Yi Cheng, Hehe Fan, Dongyun Lin et al.

The main challenge in video question answering (VideoQA) is to capture and understand the complex spatial and temporal relations between objects based on given questions. Existing graph-based methods for VideoQA usually ignore keywords in questions and employ a simple graph to aggregate features without considering relative relations between objects, which may lead to inferior performance. In this paper, we propose a Keyword-aware Relative Spatio-Temporal (KRST) graph network for VideoQA. First, to make question features aware of keywords, we employ an attention mechanism to assign high weights to keywords during question encoding. The keyword-aware question features are then used to guide video graph construction. Second, because relations are relative, we integrate the relative relation modeling to better capture the spatio-temporal dynamics among object nodes. Moreover, we disentangle the spatio-temporal reasoning into an object-level spatial graph and a frame-level temporal graph, which reduces the impact of spatial and temporal relation reasoning on each other. Extensive experiments on the TGIF-QA, MSVD-QA and MSRVTT-QA datasets demonstrate the superiority of our KRST over multiple state-of-the-art methods.

Wenjin Hou, Wei Liu, Han Hu et al.

Multimodal Large Language Models (MLLMs) have shown remarkable proficiency on general-purpose vision-language benchmarks, reaching or even exceeding human-level performance. However, these evaluations typically rely on standard in-distribution data, leaving the robustness of MLLMs largely unexamined when faced with scenarios that defy common-sense priors. To address this gap, we introduce VIA-Bench, a challenging benchmark designed to probe model performance on visual illusions and anomalies. It includes six core categories: color illusions, motion illusions, gestalt illusions, geometric and spatial illusions, general visual illusions, and visual anomalies. Through careful human-in-the-loop review, we construct over 1K high-quality question-answer pairs that require nuanced visual reasoning. Extensive evaluation of over 20 state-of-the-art MLLMs, including proprietary, open-source, and reasoning-enhanced models, uncovers significant vulnerabilities. Notably, we find that Chain-of-Thought (CoT) reasoning offers negligible robustness, often yielding ``brittle mirages'' where the model's logic collapses under illusory stimuli. Our findings reveal a fundamental divergence between machine and human perception, suggesting that resolving such perceptual bottlenecks is critical for the advancement of artificial general intelligence. The benchmark data and code will be released.

Yu Lu, Linchao Zhu, Hehe Fan et al.

Text-to-video (T2V) generation is a rapidly growing research area that aims to translate the scenes, objects, and actions within complex video text into a sequence of coherent visual frames. We present FlowZero, a novel framework that combines Large Language Models (LLMs) with image diffusion models to generate temporally-coherent videos. FlowZero uses LLMs to understand complex spatio-temporal dynamics from text, where LLMs can generate a comprehensive dynamic scene syntax (DSS) containing scene descriptions, object layouts, and background motion patterns. These elements in DSS are then used to guide the image diffusion model for video generation with smooth object motions and frame-to-frame coherence. Moreover, FlowZero incorporates an iterative self-refinement process, enhancing the alignment between the spatio-temporal layouts and the textual prompts for the videos. To enhance global coherence, we propose enriching the initial noise of each frame with motion dynamics to control the background movement and camera motion adaptively. By using spatio-temporal syntaxes to guide the diffusion process, FlowZero achieves improvement in zero-shot video synthesis, generating coherent videos with vivid motion.

Yi Cheng, Ziwei Xu, Fen Fang et al.

In this technical report, we present our findings from a study conducted on the EPIC-KITCHENS-100 Unsupervised Domain Adaptation task for Action Recognition. Our research focuses on the innovative application of a differentiable logic loss in the training to leverage the co-occurrence relations between verb and noun, as well as the pre-trained Large Language Models (LLMs) to generate the logic rules for the adaptation to unseen action labels. Specifically, the model's predictions are treated as the truth assignment of a co-occurrence logic formula to compute the logic loss, which measures the consistency between the predictions and the logic constraints. By using the verb-noun co-occurrence matrix generated from the dataset, we observe a moderate improvement in model performance compared to our baseline framework. To further enhance the model's adaptability to novel action labels, we experiment with rules generated using GPT-3.5, which leads to a slight decrease in performance. These findings shed light on the potential and challenges of incorporating differentiable logic and LLMs for knowledge extraction in unsupervised domain adaptation for action recognition. Our final submission (entitled `NS-LLM') achieved the first place in terms of top-1 action recognition accuracy.

Xiaoxiao Sheng, Zhiqiang Shen, Gang Xiao et al.

We propose a unified point cloud video self-supervised learning framework for object-centric and scene-centric data. Previous methods commonly conduct representation learning at the clip or frame level and cannot well capture fine-grained semantics. Instead of contrasting the representations of clips or frames, in this paper, we propose a unified self-supervised framework by conducting contrastive learning at the point level. Moreover, we introduce a new pretext task by achieving semantic alignment of superpoints, which further facilitates the representations to capture semantic cues at multiple scales. In addition, due to the high redundancy in the temporal dimension of dynamic point clouds, directly conducting contrastive learning at the point level usually leads to massive undesired negatives and insufficient modeling of positive representations. To remedy this, we propose a selection strategy to retain proper negatives and make use of high-similarity samples from other instances as positive supplements. Extensive experiments show that our method outperforms supervised counterparts on a wide range of downstream tasks and demonstrates the superior transferability of the learned representations.

Ming Li, Xiangyu Xu, Hehe Fan et al.

Existing methods of privacy-preserving action recognition (PPAR) mainly focus on frame-level (spatial) privacy removal through 2D CNNs. Unfortunately, they have two major drawbacks. First, they may compromise temporal dynamics in input videos, which are critical for accurate action recognition. Second, they are vulnerable to practical attacking scenarios where attackers probe for privacy from an entire video rather than individual frames. To address these issues, we propose a novel framework STPrivacy to perform video-level PPAR. For the first time, we introduce vision Transformers into PPAR by treating a video as a tubelet sequence, and accordingly design two complementary mechanisms, i.e., sparsification and anonymization, to remove privacy from a spatio-temporal perspective. In specific, our privacy sparsification mechanism applies adaptive token selection to abandon action-irrelevant tubelets. Then, our anonymization mechanism implicitly manipulates the remaining action-tubelets to erase privacy in the embedding space through adversarial learning. These mechanisms provide significant advantages in terms of privacy preservation for human eyes and action-privacy trade-off adjustment during deployment. We additionally contribute the first two large-scale PPAR benchmarks, VP-HMDB51 and VP-UCF101, to the community. Extensive evaluations on them, as well as two other tasks, validate the effectiveness and generalization capability of our framework.

Yue Zhang, Hehe Fan, Yi Yang et al.

In this technical report, we present our findings from the research conducted on the Human-Object Interaction 4D (HOI4D) dataset for egocentric action segmentation task. As a relatively novel research area, point cloud video methods might not be good at temporal modeling, especially for long point cloud videos (\eg, 150 frames). In contrast, traditional video understanding methods have been well developed. Their effectiveness on temporal modeling has been widely verified on many large scale video datasets. Therefore, we convert point cloud videos into depth videos and employ traditional video modeling methods to improve 4D action segmentation. By ensembling depth and point cloud video methods, the accuracy is significantly improved. The proposed method, named Mixture of Depth and Point cloud video experts (DPMix), achieved the first place in the 4D Action Segmentation Track of the HOI4D Challenge 2023.

Wenjie Zhuo, Fan Ma, Hehe Fan et al.

This paper presents Invariant Score Distillation (ISD), a novel method for high-fidelity text-to-3D generation. ISD aims to tackle the over-saturation and over-smoothing problems in Score Distillation Sampling (SDS). In this paper, SDS is decoupled into a weighted sum of two components: the reconstruction term and the classifier-free guidance term. We experimentally found that over-saturation stems from the large classifier-free guidance scale and over-smoothing comes from the reconstruction term. To overcome these problems, ISD utilizes an invariant score term derived from DDIM sampling to replace the reconstruction term in SDS. This operation allows the utilization of a medium classifier-free guidance scale and mitigates the reconstruction-related errors, thus preventing the over-smoothing and over-saturation of results. Extensive experiments demonstrate that our method greatly enhances SDS and produces realistic 3D objects through single-stage optimization.

Guoliang Chen, Fei Wang, Kun Li et al.

In this paper, we briefly introduce the solution developed by our team, HFUT-VUT, for the track of Micro-gesture Classification in the MiGA challenge at IJCAI 2024. The task of micro-gesture classification task involves recognizing the category of a given video clip, which focuses on more fine-grained and subtle body movements compared to typical action recognition tasks. Given the inherent complexity of micro-gesture recognition, which includes large intra-class variability and minimal inter-class differences, we utilize two innovative modules, i.e., the cross-modal fusion module and prototypical refinement module, to improve the discriminative ability of MG features, thereby improving the classification accuracy. Our solution achieved significant success, ranking 1st in the track of Micro-gesture Classification. We surpassed the performance of last year's leading team by a substantial margin, improving Top-1 accuracy by 6.13%.

Yuxuan Hou, Hehe Fan, Jianrong Zhang et al.

The automatic synthesis of analog circuits presents significant challenges. Most existing approaches formulate the problem as a single-objective optimization task, overlooking that design specifications for a given circuit type vary widely across applications. To address this, we introduce specification-conditioned analog circuit generation, a task that directly generates analog circuits based on target specifications. The motivation is to leverage existing well-designed circuits to improve automation in analog circuit design. Specifically, we propose CktGen, a simple yet effective variational autoencoder that maps discretized specifications and circuits into a joint latent space and reconstructs the circuit from that latent vector. Notably, as a single specification may correspond to multiple valid circuits, naively fusing specification information into the generative model does not capture these one-to-many relationships. To address this, we decouple the encoding of circuits and specifications and align their mapped latent space. Then, we employ contrastive training with a filter mask to maximize differences between encoded circuits and specifications. Furthermore, classifier guidance along with latent feature alignment promotes the clustering of circuits sharing the same specification, avoiding model collapse into trivial one-to-one mappings. By canonicalizing the latent space with respect to specifications, we can search for an optimal circuit that meets valid target specifications. We conduct comprehensive experiments on the open circuit benchmark and introduce metrics to evaluate cross-model consistency. Experimental results demonstrate that CktGen achieves substantial improvements over state-of-the-art methods.

Ruisi Zhao, Haoren Zheng, Zongxin Yang et al.

Rigged 3D assets are fundamental to 3D deformation and animation. However, existing 3D generation methods face challenges in generating animatable geometry, while rigging techniques lack fine-grained structural control over skeleton creation. To address these limitations, we introduce Stroke3D, a novel framework that directly generates rigged meshes from user inputs: 2D drawn strokes and a descriptive text prompt. Our approach pioneers a two-stage pipeline that separates the generation into: 1) Controllable Skeleton Generation, we employ the Skeletal Graph VAE (Sk-VAE) to encode the skeleton's graph structure into a latent space, where the Skeletal Graph DiT (Sk-DiT) generates a skeletal embedding. The generation process is conditioned on both the text for semantics and the 2D strokes for explicit structural control, with the VAE's decoder reconstructing the final high-quality 3D skeleton; and 2) Enhanced Mesh Synthesis via TextuRig and SKA-DPO, where we then synthesize a textured mesh conditioned on the generated skeleton. For this stage, we first enhance an existing skeleton-to-mesh model by augmenting its training data with TextuRig: a dataset of textured and rigged meshes with captions, curated from Objaverse-XL. Additionally, we employ a preference optimization strategy, SKA-DPO, guided by a skeleton-mesh alignment score, to further improve geometric fidelity. Together, our framework enables a more intuitive workflow for creating ready to animate 3D content. To the best of our knowledge, our work is the first to generate rigged 3D meshes conditioned on user-drawn 2D strokes. Extensive experiments demonstrate that Stroke3D produces plausible skeletons and high-quality meshes.

Hang Du, Sicheng Zhang, Binzhu Xie et al.

Video anomaly understanding (VAU) aims to automatically comprehend unusual occurrences in videos, thereby enabling various applications such as traffic surveillance and industrial manufacturing. While existing VAU benchmarks primarily concentrate on anomaly detection and localization, our focus is on more practicality, prompting us to raise the following crucial questions: "what anomaly occurred?", "why did it happen?", and "how severe is this abnormal event?". In pursuit of these answers, we present a comprehensive benchmark for Causation Understanding of Video Anomaly (CUVA). Specifically, each instance of the proposed benchmark involves three sets of human annotations to indicate the "what", "why" and "how" of an anomaly, including 1) anomaly type, start and end times, and event descriptions, 2) natural language explanations for the cause of an anomaly, and 3) free text reflecting the effect of the abnormality. In addition, we also introduce MMEval, a novel evaluation metric designed to better align with human preferences for CUVA, facilitating the measurement of existing LLMs in comprehending the underlying cause and corresponding effect of video anomalies. Finally, we propose a novel prompt-based method that can serve as a baseline approach for the challenging CUVA. We conduct extensive experiments to show the superiority of our evaluation metric and the prompt-based approach. Our code and dataset are available at https://github.com/fesvhtr/CUVA.

Kun Li, Dan Guo, Guoliang Chen et al.

Micro-Action Recognition (MAR) has gained increasing attention due to its crucial role as a form of non-verbal communication in social interactions, with promising potential for applications in human communication and emotion analysis. However, current approaches often overlook the inherent ambiguity in micro-actions, which arises from the wide category range and subtle visual differences between categories. This oversight hampers the accuracy of micro-action recognition. In this paper, we propose a novel Prototypical Calibrating Ambiguous Network (PCAN) to unleash and mitigate the ambiguity of MAR. Firstly, we employ a hierarchical action-tree to identify the ambiguous sample, categorizing them into distinct sets of ambiguous samples of false negatives and false positives, considering both body- and action-level categories. Secondly, we implement an ambiguous contrastive refinement module to calibrate these ambiguous samples by regulating the distance between ambiguous samples and their corresponding prototypes. This calibration process aims to pull false negative (FN) samples closer to their respective prototypes and push false positive (FP) samples apart from their affiliated prototypes. In addition, we propose a new prototypical diversity amplification loss to strengthen the model's capacity by amplifying the differences between different prototypes. Finally, we propose a prototype-guided rectification to rectify prediction by incorporating the representability of prototypes. Extensive experiments conducted on the benchmark dataset demonstrate the superior performance of our method compared to existing approaches. The code is available at https://github.com/kunli-cs/PCAN.

Zhenglin Zhou, Fan Ma, Hehe Fan et al.

Creating digital avatars from textual prompts has long been a desirable yet challenging task. Despite the promising results achieved with 2D diffusion priors, current methods struggle to create high-quality and consistent animated avatars efficiently. Previous animatable head models like FLAME have difficulty in accurately representing detailed texture and geometry. Additionally, high-quality 3D static representations face challenges in semantically driving with dynamic priors. In this paper, we introduce \textbf{HeadStudio}, a novel framework that utilizes 3D Gaussian splatting to generate realistic and animatable avatars from text prompts. Firstly, we associate 3D Gaussians with animatable head prior model, facilitating semantic animation on high-quality 3D representations. To ensure consistent animation, we further enhance the optimization from initialization, distillation, and regularization to jointly learn the shape, texture, and animation. Extensive experiments demonstrate the efficacy of HeadStudio in generating animatable avatars from textual prompts, exhibiting appealing appearances. The avatars are capable of rendering high-quality real-time ($\geq 40$ fps) novel views at a resolution of 1024. Moreover, These avatars can be smoothly driven by real-world speech and video. We hope that HeadStudio can enhance digital avatar creation and gain popularity in the community. Code is at: https://github.com/ZhenglinZhou/HeadStudio.

Xiangpeng Yang, Linchao Zhu, Hehe Fan et al.

Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available at https://knightyxp.github.io/VideoGrain_project_page/

Zhenglin Zhou, Xiaobo Xia, Fan Ma et al.

Text-to-3D generation automates 3D content creation from textual descriptions, which offers transformative potential across various fields. However, existing methods often struggle to align generated content with human preferences, limiting their applicability and flexibility. To address these limitations, in this paper, we propose DreamDPO, an optimization-based framework that integrates human preferences into the 3D generation process, through direct preference optimization. Practically, DreamDPO first constructs pairwise examples, then compare their alignment with human preferences using reward or large multimodal models, and lastly optimizes the 3D representation with a preference-driven loss function. By leveraging pairwise comparison to reflect preferences, DreamDPO reduces reliance on precise pointwise quality evaluations while enabling fine-grained controllability through preference-guided optimization. Experiments demonstrate that DreamDPO achieves competitive results, and provides higher-quality and more controllable 3D content compared to existing methods. The code and models will be open-sourced.

Hang Du, Guoshun Nan, Sicheng Zhang et al.

Multimodal Sarcasm Understanding (MSU) has a wide range of applications in the news field such as public opinion analysis and forgery detection. However, existing MSU benchmarks and approaches usually focus on sentence-level MSU. In document-level news, sarcasm clues are sparse or small and are often concealed in long text. Moreover, compared to sentence-level comments like tweets, which mainly focus on only a few trends or hot topics (e.g., sports events), content in the news is considerably diverse. Models created for sentence-level MSU may fail to capture sarcasm clues in document-level news. To fill this gap, we present a comprehensive benchmark for Document-level Multimodal Sarcasm Understanding (DocMSU). Our dataset contains 102,588 pieces of news with text-image pairs, covering 9 diverse topics such as health, business, etc. The proposed large-scale and diverse DocMSU significantly facilitates the research of document-level MSU in real-world scenarios. To take on the new challenges posed by DocMSU, we introduce a fine-grained sarcasm comprehension method to properly align the pixel-level image features with word-level textual features in documents. Experiments demonstrate the effectiveness of our method, showing that it can serve as a baseline approach to the challenging DocMSU. Our code and dataset are available at https://github.com/Dulpy/DocMSU.

Zhenglin Zhou, Fan Ma, Xiaobo Xia et al.

We explore inference-time scaling in text-guided 3D diffusion models to enhance generative quality without additional training. To this end, we introduce ITS3D, a framework that formulates the task as an optimization problem to identify the most effective Gaussian noise input. The framework is driven by a verifier-guided search algorithm, where the search algorithm iteratively refines noise candidates based on verifier feedback. To address the inherent challenges of 3D generation, we introduce three techniques for improved stability, efficiency, and exploration capability. 1) Gaussian normalization is applied to stabilize the search process. It corrects distribution shifts when noise candidates deviate from a standard Gaussian distribution during iterative updates. 2) The high-dimensional nature of the 3D search space increases computational complexity. To mitigate this, a singular value decomposition-based compression technique is employed to reduce dimensionality while preserving effective search directions. 3) To further prevent convergence to suboptimal local minima, a singular space reset mechanism dynamically updates the search space based on diversity measures. Extensive experiments demonstrate that ITS3D enhances text-to-3D generation quality, which shows the potential of computationally efficient search methods in generative processes. The source code is available at https://github.com/ZhenglinZhou/ITS3D.

Zhenglin Zhou, Fan Ma, Chengzhuo Gui et al.

Training-free 3D editing aims to modify 3D shapes based on human instructions without model finetuning. It plays a crucial role in 3D content creation. However, existing approaches often struggle to produce strong or geometrically stable edits, largely due to inconsistent latent anchors introduced by timestep-dependent noise during diffusion sampling. To address these limitations, we introduce AnchorFlow, which is built upon the principle of latent anchor consistency. Specifically, AnchorFlow establishes a global latent anchor shared between the source and target trajectories, and enforces coherence using a relaxed anchor-alignment loss together with an anchor-aligned update rule. This design ensures that transformations remain stable and semantically faithful throughout the editing process. By stabilizing the latent reference space, AnchorFlow enables more pronounced semantic modifications. Moreover, AnchorFlow is mask-free. Without mask supervision, it effectively preserves geometric fidelity. Experiments on the Eval3DEdit benchmark show that AnchorFlow consistently delivers semantically aligned and structurally robust edits across diverse editing types. Code is at https://github.com/ZhenglinZhou/AnchorFlow.

Jianrong Zhang, Hehe Fan, Yi Yang

Human motions are compositional: complex behaviors can be described as combinations of simpler primitives. However, existing approaches primarily focus on forward modeling, e.g., learning holistic mappings from text to motion or composing a complex motion from a set of motion concepts. In this paper, we consider the inverse perspective: decomposing a holistic motion into semantically meaningful sub-components. We propose DeMoGen, a compositional training paradigm for decompositional learning that employs an energy-based diffusion model. This energy formulation directly captures the composed distribution of multiple motion concepts, enabling the model to discover them without relying on ground-truth motions for individual concepts. Within this paradigm, we introduce three training variants to encourage a decompositional understanding of motion: 1. DeMoGen-Exp explicitly trains on decomposed text prompts; 2. DeMoGen-OSS performs orthogonal self-supervised decomposition; 3. DeMoGen-SC enforces semantic consistency between original and decomposed text embeddings. These variants enable our approach to disentangle reusable motion primitives from complex motion sequences. We also demonstrate that the decomposed motion concepts can be flexibly recombined to generate diverse and novel motions, generalizing beyond the training distribution. Additionally, we construct a text-decomposed dataset to support compositional training, serving as an extended resource to facilitate text-to-motion generation and motion composition.

Hehe Fan, Yi Yang

In this paper, we introduce a Point Recurrent Neural Network (PointRNN) for moving point cloud processing. At each time step, PointRNN takes point coordinates $\boldsymbol{P} \in \mathbb{R}^{n \times 3}$ and point features $\boldsymbol{X} \in \mathbb{R}^{n \times d}$ as input ($n$ and $d$ denote the number of points and the number of feature channels, respectively). The state of PointRNN is composed of point coordinates $\boldsymbol{P}$ and point states $\boldsymbol{S} \in \mathbb{R}^{n \times d'}$ ($d'$ denotes the number of state channels). Similarly, the output of PointRNN is composed of $\boldsymbol{P}$ and new point features $\boldsymbol{Y} \in \mathbb{R}^{n \times d''}$ ($d''$ denotes the number of new feature channels). Since point clouds are orderless, point features and states from two time steps can not be directly operated. Therefore, a point-based spatiotemporally-local correlation is adopted to aggregate point features and states according to point coordinates. We further propose two variants of PointRNN, i.e., Point Gated Recurrent Unit (PointGRU) and Point Long Short-Term Memory (PointLSTM). We apply PointRNN, PointGRU and PointLSTM to moving point cloud prediction, which aims to predict the future trajectories of points in a set given their history movements. Experimental results show that PointRNN, PointGRU and PointLSTM are able to produce correct predictions on both synthetic and real-world datasets, demonstrating their ability to model point cloud sequences. The code has been released at \url{https://github.com/hehefan/PointRNN}.

Yuhang Ding, Hehe Fan, Mingliang Xu et al.

Due to domain bias, directly deploying a deep person re-identification (re-ID) model trained on one dataset often achieves considerably poor accuracy on another dataset. In this paper, we propose an Adaptive Exploration (AE) method to address the domain-shift problem for re-ID in an unsupervised manner. Specifically, in the target domain, the re-ID model is inducted to 1) maximize distances between all person images and 2) minimize distances between similar person images. In the first case, by treating each person image as an individual class, a non-parametric classifier with a feature memory is exploited to encourage person images to move far away from each other. In the second case, according to a similarity threshold, our method adaptively selects neighborhoods for each person image in the feature space. By treating these similar person images as the same class, the non-parametric classifier forces them to stay closer. However, a problem of the adaptive selection is that, when an image has too many neighborhoods, it is more likely to attract other images as its neighborhoods. As a result, a minority of images may select a large number of neighborhoods while a majority of images have only a few neighborhoods. To address this issue, we additionally integrate a balance strategy into the adaptive selection. We evaluate our methods with two protocols. The first one is called "target-only re-ID", in which only the unlabeled target data is used for training. The second one is called "domain adaptive re-ID", in which both the source data and the target data are used during training. Experimental results on large-scale re-ID datasets demonstrate the effectiveness of our method. Our code has been released at https://github.com/dyh127/Adaptive-Exploration-for-Unsupervised-Person-Re-Identification.

Xiangpeng Yang, Linchao Zhu, Hehe Fan et al.

Current diffusion-based video editing primarily focuses on local editing (\textit{e.g.,} object/background editing) or global style editing by utilizing various dense correspondences. However, these methods often fail to accurately edit the foreground and background simultaneously while preserving the original layout. We find that the crux of the issue stems from the imprecise distribution of attention weights across designated regions, including inaccurate text-to-attribute control and attention leakage. To tackle this issue, we introduce EVA, a \textbf{zero-shot} and \textbf{multi-attribute} video editing framework tailored for human-centric videos with complex motions. We incorporate a Spatial-Temporal Layout-Guided Attention mechanism that leverages the intrinsic positive and negative correspondences of cross-frame diffusion features. To avoid attention leakage, we utilize these correspondences to boost the attention scores of tokens within the same attribute across all video frames while limiting interactions between tokens of different attributes in the self-attention layer. For precise text-to-attribute manipulation, we use discrete text embeddings focused on specific layout areas within the cross-attention layer. Benefiting from the precise attention weight distribution, EVA can be easily generalized to multi-object editing scenarios and achieves accurate identity mapping. Extensive experiments demonstrate EVA achieves state-of-the-art results in real-world scenarios. Full results are provided at https://knightyxp.github.io/EVA/

Xiaobo Hu, Youfang Lin, HeHe Fan et al.

Given an object of interest, visual navigation aims to reach the object's location based on a sequence of partial observations. To this end, an agent needs to 1) learn a piece of certain knowledge about the relations of object categories in the world during training and 2) look for the target object based on the pre-learned object category relations and its moving trajectory in the current unseen environment. In this paper, we propose a Category Relation Graph (CRG) to learn the knowledge of object category layout relations and a Temporal-Spatial-Region (TSR) attention architecture to perceive the long-term spatial-temporal dependencies of objects helping the navigation. We learn prior knowledge of object layout, establishing a category relationship graph to deduce the positions of specific objects. Subsequently, we introduced TSR to capture the relationships of objects in temporal, spatial, and regions within the observation trajectories. Specifically, we propose a Temporal attention module (T) to model the temporal structure of the observation sequence, which implicitly encodes the historical moving or trajectory information. Then, a Spatial attention module (S) is used to uncover the spatial context of the current observation objects based on the category relation graph and past observations. Last, a Region attention module (R) shifts the attention to the target-relevant region. Based on the visual representation extracted by our method, the agent can better perceive the environment and easily learn superior navigation policy. Experiments on AI2-THOR demonstrate our CRG-TSR method significantly outperforms existing methods regarding both effectiveness and efficiency. The code has been included in the supplementary material and will be publicly available.

Wei Li, Hehe Fan, Yongkang Wong et al.

Recent advancements in image understanding have benefited from the extensive use of web image-text pairs. However, video understanding remains a challenge despite the availability of substantial web video-text data. This difficulty primarily arises from the inherent complexity of videos and the inefficient language supervision in recent web-collected video-text datasets. In this paper, we introduce Text-Only Pre-Alignment (TOPA), a novel approach to extend large language models (LLMs) for video understanding, without the need for pre-training on real video data. Specifically, we first employ an advanced LLM to automatically generate Textual Videos comprising continuous textual frames, along with corresponding annotations to simulate real video-text data. Then, these annotated textual videos are used to pre-align a language-only LLM with the video modality. To bridge the gap between textual and real videos, we employ the CLIP model as the feature extractor to align image and text modalities. During text-only pre-alignment, the continuous textual frames, encoded as a sequence of CLIP text features, are analogous to continuous CLIP image features, thus aligning the LLM with real video representation. Extensive experiments, including zero-shot evaluation and finetuning on various video understanding tasks, demonstrate that TOPA is an effective and efficient framework for aligning video content with LLMs. In particular, without training on any video data, the TOPA-Llama2-13B model achieves a Top-1 accuracy of 51.0% on the challenging long-form video understanding benchmark, Egoschema. This performance surpasses previous video-text pre-training approaches and proves competitive with recent GPT-3.5-based video agents.

Ruijie Quan, Wenguan Wang, Fan Ma et al.

Protein representation learning is a challenging task that aims to capture the structure and function of proteins from their amino acid sequences. Previous methods largely ignored the fact that not all amino acids are equally important for protein folding and activity. In this article, we propose a neural clustering framework that can automatically discover the critical components of a protein by considering both its primary and tertiary structure information. Our framework treats a protein as a graph, where each node represents an amino acid and each edge represents a spatial or sequential connection between amino acids. We then apply an iterative clustering strategy to group the nodes into clusters based on their 1D and 3D positions and assign scores to each cluster. We select the highest-scoring clusters and use their medoid nodes for the next iteration of clustering, until we obtain a hierarchical and informative representation of the protein. We evaluate on four protein-related tasks: protein fold classification, enzyme reaction classification, gene ontology term prediction, and enzyme commission number prediction. Experimental results demonstrate that our method achieves state-of-the-art performance.

Yuze Hao, Jianrong Zhang, Tao Zhuo et al.

Hands are the main medium when people interact with the world. Generating proper 3D motion for hand-object interaction is vital for applications such as virtual reality and robotics. Although grasp tracking or object manipulation synthesis can produce coarse hand motion, this kind of motion is inevitably noisy and full of jitter. To address this problem, we propose a data-driven method for coarse motion refinement. First, we design a hand-centric representation to describe the dynamic spatial-temporal relation between hands and objects. Compared to the object-centric representation, our hand-centric representation is straightforward and does not require an ambiguous projection process that converts object-based prediction into hand motion. Second, to capture the dynamic clues of hand-object interaction, we propose a new architecture that models the spatial and temporal structure in a hierarchical manner. Extensive experiments demonstrate that our method outperforms previous methods by a noticeable margin.

Yue Zhang, Hehe Fan, Yi Yang

To bridge the gap between vision and language modalities, Multimodal Large Language Models (MLLMs) usually learn an adapter that converts visual inputs to understandable tokens for Large Language Models (LLMs). However, most adapters generate consistent visual tokens, regardless of the specific objects of interest mentioned in the prompt. Since these adapters distribute equal attention to every detail in the image and focus on the entire scene, they may increase the cognitive load for LLMs, particularly when processing complex scenes. To alleviate this problem, we propose prompt-aware adapters. These adapters are designed with the capability to dynamically embed visual inputs based on the specific focus of the prompt. Specifically, prompt-aware adapters utilize both global and local textual features to capture the most relevant visual clues from the prompt at both coarse and fine granularity levels. This approach significantly enhances the ability of LLMs to understand and interpret visual content. Experiments on various visual question answering tasks, such as counting and position reasoning, demonstrate the effectiveness of prompt-aware adapters.

Mingwei Li, Pu Pang, Hehe Fan et al.

Reconstructing transparent surfaces is essential for tasks such as robotic manipulation in labs, yet it poses a significant challenge for 3D reconstruction techniques like 3D Gaussian Splatting (3DGS). These methods often encounter a transparency-depth dilemma, where the pursuit of photorealistic rendering through standard $α$-blending undermines geometric precision, resulting in considerable depth estimation errors for transparent materials. To address this issue, we introduce Transparent Surface Gaussian Splatting (TSGS), a new framework that separates geometry learning from appearance refinement. In the geometry learning stage, TSGS focuses on geometry by using specular-suppressed inputs to accurately represent surfaces. In the second stage, TSGS improves visual fidelity through anisotropic specular modeling, crucially maintaining the established opacity to ensure geometric accuracy. To enhance depth inference, TSGS employs a first-surface depth extraction method. This technique uses a sliding window over $α$-blending weights to pinpoint the most likely surface location and calculates a robust weighted average depth. To evaluate the transparent surface reconstruction task under realistic conditions, we collect a TransLab dataset that includes complex transparent laboratory glassware. Extensive experiments on TransLab show that TSGS achieves accurate geometric reconstruction and realistic rendering of transparent objects simultaneously within the efficient 3DGS framework. Specifically, TSGS significantly surpasses current leading methods, achieving a 37.3% reduction in chamfer distance and an 8.0% improvement in F1 score compared to the top baseline. The code and dataset are available at https://longxiang-ai.github.io/TSGS/.

Jianrong Zhang, Hehe Fan, Yi Yang

Diffusion models, particularly latent diffusion models, have demonstrated remarkable success in text-driven human motion generation. However, it remains challenging for latent diffusion models to effectively compose multiple semantic concepts into a single, coherent motion sequence. To address this issue, we propose EnergyMoGen, which includes two spectrums of Energy-Based Models: (1) We interpret the diffusion model as a latent-aware energy-based model that generates motions by composing a set of diffusion models in latent space; (2) We introduce a semantic-aware energy model based on cross-attention, which enables semantic composition and adaptive gradient descent for text embeddings. To overcome the challenges of semantic inconsistency and motion distortion across these two spectrums, we introduce Synergistic Energy Fusion. This design allows the motion latent diffusion model to synthesize high-quality, complex motions by combining multiple energy terms corresponding to textual descriptions. Experiments show that our approach outperforms existing state-of-the-art models on various motion generation tasks, including text-to-motion generation, compositional motion generation, and multi-concept motion generation. Additionally, we demonstrate that our method can be used to extend motion datasets and improve the text-to-motion task.

Wenjie Zhuo, Fan Ma, Hehe Fan

We present InfiniDreamer, a novel framework for arbitrarily long human motion generation. InfiniDreamer addresses the limitations of current motion generation methods, which are typically restricted to short sequences due to the lack of long motion training data. To achieve this, we first generate sub-motions corresponding to each textual description and then assemble them into a coarse, extended sequence using randomly initialized transition segments. We then introduce an optimization-based method called Segment Score Distillation (SSD) to refine the entire long motion sequence. SSD is designed to utilize an existing motion prior, which is trained only on short clips, in a training-free manner. Specifically, SSD iteratively refines overlapping short segments sampled from the coarsely extended long motion sequence, progressively aligning them with the pre-trained motion diffusion prior. This process ensures local coherence within each segment, while the refined transitions between segments maintain global consistency across the entire sequence. Extensive qualitative and quantitative experiments validate the superiority of our framework, showcasing its ability to generate coherent, contextually aware motion sequences of arbitrary length.

Sai Wang, Fan Ma, Xinyi Li et al.

Recent advancements in LLMs have accelerated the development of dialogue generation across text and images, yet video-based dialogue generation remains underexplored and presents unique challenges. In this paper, we introduce Theme-aware Video Dialogue Crafting (TVDC), a novel task aimed at generating new dialogues that align with video content and adhere to user-specified themes. We propose TV-Dialogue, a novel multi-modal agent framework that ensures both theme alignment (i.e., the dialogue revolves around the theme) and visual consistency (i.e., the dialogue matches the emotions and behaviors of characters in the video) by enabling real-time immersive interactions among video characters, thereby accurately understanding the video content and generating new dialogue that aligns with the given themes. To assess the generated dialogues, we present a multi-granularity evaluation benchmark with high accuracy, interpretability and reliability, demonstrating the effectiveness of TV-Dialogue on self-collected dataset over directly using existing LLMs. Extensive experiments reveal that TV-Dialogue can generate dialogues for videos of any length and any theme in a zero-shot manner without training. Our findings underscore the potential of TV-Dialogue for various applications, such as video re-creation, film dubbing and its use in downstream multimodal tasks.

Xiaobo Hu, Youfang Lin, Yue Liu et al.

Visual reinforcement learning has proven effective in solving control tasks with high-dimensional observations. However, extracting reliable and generalizable representations from vision-based observations remains a central challenge. Inspired by the human thought process, when the representation extracted from the observation can predict the future and trace history, the representation is reliable and accurate in comprehending the environment. Based on this concept, we introduce a Bidirectional Transition (BiT) model, which leverages the ability to bidirectionally predict environmental transitions both forward and backward to extract reliable representations. Our model demonstrates competitive generalization performance and sample efficiency on two settings of the DeepMind Control suite. Additionally, we utilize robotic manipulation and CARLA simulators to demonstrate the wide applicability of our method.

Xindan Zhang, Weilong Yan, Yufei Shi et al.

Point clouds provide a compact and expressive representation of 3D objects, and have recently been integrated into multimodal large language models (MLLMs). However, existing methods primarily focus on static objects, while understanding dynamic point cloud sequences remains largely unexplored. This limitation is mainly caused by the lack of large-scale cross-modal datasets and the difficulty of modeling motions in spatio-temporal contexts. To bridge this gap, we present 4DPC$^2$hat, the first MLLM tailored for dynamic point cloud understanding. To this end, we construct a large-scale cross-modal dataset 4DPC$^2$hat-200K via a meticulous two-stage pipeline consisting of topology-consistent 4D point construction and two-level captioning. The dataset contains over 44K dynamic object sequences, 700K point cloud frames, and 200K curated question-answer (QA) pairs, supporting inquiries about counting, temporal relationship, action, spatial relationship, and appearance. At the core of the framework, we introduce a Mamba-enhanced temporal reasoning MLLM to capture long-range dependencies and dynamic patterns among a point cloud sequence. Furthermore, we propose a failure-aware bootstrapping learning strategy that iteratively identifies model deficiencies and generates targeted QA supervision to continuously strengthen corresponding reasoning capabilities. Extensive experiments demonstrate that our 4DPC$^2$hat significantly improves action understanding and temporal reasoning compared with existing models, establishing a strong foundation for 4D dynamic point cloud understanding.

Ying Shu, Pujian Zhan, Huiqi Yang et al.

Both fine-grained discriminative details and global semantic features can contribute to solving person re-identification challenges, such as occlusion and pose variations. Vision foundation models (\textit{e.g.}, DINO) excel at mining local textures, and vision-language models (\textit{e.g.}, CLIP) capture strong global semantic difference. Existing methods predominantly rely on a single paradigm, neglecting the potential benefits of their integration. In this paper, we analyze the complementary roles of these two architectures and propose a framework to synergize their strengths by a \textbf{D}ual-\textbf{R}egularized Bidirectional \textbf{Transformer} (\textbf{DRFormer}). The dual-regularization mechanism ensures diverse feature extraction and achieves a better balance in the contributions of the two models. Extensive experiments on five benchmarks show that our method effectively harmonizes local and global representations, achieving competitive performance against state-of-the-art methods.

Haoyuan Li, Rui Liu, Hehe Fan et al.

Enabling agents to understand and interact with complex 3D scenes is a fundamental challenge for embodied artificial intelligence systems. While Multimodal Large Language Models (MLLMs) have achieved significant progress in 2D image understanding, extending such capabilities to 3D scenes remains difficult: 1) 3D environment involves richer concepts such as spatial relationships, affordances, physics, layout, and so on, 2) the absence of large-scale 3D vision-language datasets has posed a significant obstacle. In this paper, we introduce Text-Scene, a framework that automatically parses 3D scenes into textual descriptions for scene understanding. Given a 3D scene, our model identifies object attributes and spatial relationships, and then generates a coherent summary of the whole scene, bridging the gap between 3D observation and language without requiring human-in-the-loop intervention. By leveraging both geometric analysis and MLLMs, Text-Scene produces descriptions that are accurate, detailed, and human-interpretable, capturing object-level details and global-level context. Experimental results on benchmarks demonstrate that our textual parses can faithfully represent 3D scenes and benefit downstream tasks. To evaluate the reasoning capability of MLLMs, we present InPlan3D, a comprehensive benchmark for 3D task planning, consisting of 3174 long-term planning tasks across 636 indoor scenes. We emphasize clarity and accessibility in our approach, aiming to make 3D scene content understandable through language. Code and datasets will be released.

Yixiao Zhou, Ziyu Zhao, Dongzhou Cheng et al.

Sparse Mixture-of-Experts (SMoE) architectures are widely used in large language models (LLMs) due to their computational efficiency. However, though only a few experts are activated for each token, SMoE still requires loading all expert parameters, leading to high memory usage and challenges in deployment. Previous work has tried to reduce the overhead by pruning and merging experts, but primarily focused on expert-level operations, leaving neuron-level structure underexplored. We propose DERN (Dropping Experts, Recombining Neurons), a task-agnostic and retraining-free framework for expert pruning and reconstruction. We observe that experts are often misaligned and contain semantic conflicts at the neuron level, which poses challenges for direct merging. To solve this, DERN works in three steps: it first prunes redundant experts using router statistics; then it decomposes them into neuron-level expert segments, assigning each segment to its most compatible retained expert; and finally, it merges segments within each retained expert to build a compact representation. Experiments on Mixtral, Qwen, and DeepSeek SMoE models show that DERN improves performance by more than 5% on commonsense reasoning and MMLU benchmarks under 50% expert sparsity, without extra training. It also greatly reduces the number of experts and memory usage, making SMoE LLMs easier to deploy in practice.

Qianyu Feng, Yu Wu, Hehe Fan et al.

Image captioning, a challenging task where the machine automatically describes an image by sentences, has drawn significant attention in recent years. Despite the remarkable improvements of recent approaches, however, these methods are built upon a large set of training image-sentence pairs. The expensive labor efforts hence limit the captioning model to describe the wider world. In this paper, we present a novel network structure, Cascaded Revision Network, which aims at relieving the problem by equipping the model with out-of-domain knowledge. CRN first tries its best to describe an image using the existing vocabulary from in-domain knowledge. Due to the lack of out-of-domain knowledge, the caption may be inaccurate or include ambiguous words for the image with unknown (novel) objects. We propose to re-edit the primary captioning sentence by a series of cascaded operations. We introduce a perplexity predictor to find out which words are most likely to be inaccurate given the input image. Thereafter, we utilize external knowledge from a pre-trained object detection model and select more accurate words from detection results by the visual matching module. In the last step, we design a semantic matching module to ensure that the novel object is fit in the right position. By this novel cascaded captioning-revising mechanism, CRN can accurately describe images with unseen objects. We validate the proposed method with state-of-the-art performance on the held-out MSCOCO dataset as well as scale to ImageNet, demonstrating the effectiveness of this method.

Qianyu Feng, Guoliang Kang, Hehe Fan et al.

Open set domain adaptation aims to diminish the domain shift across domains, with partially shared classes. There exist unknown target samples out of the knowledge of source domain. Compared to the close set setting, how to separate the unknown (unshared) class from the known (shared) ones plays a key role. Whereas, previous methods did not emphasize the semantic structure of the open set data, which may introduce bias into the domain alignment and confuse the classifier around the decision boundary. In this paper, we exploit the semantic structure of open set data from two aspects: 1) Semantic Categorical Alignment, which aims to achieve good separability of target known classes by categorically aligning the centroid of target with the source. 2)Semantic Contrastive Mapping, which aims to push the unknown class away from the decision boundary. Empirically, we demonstrate that our method performs favourably against the state-of-the-art methods on representative benchmarks, e.g. Digit datasets and Office-31 datasets.

Hehe Fan, Linchao Zhu, Yi Yang

Predicting future frames in videos has become a promising direction of research for both computer vision and robot learning communities. The core of this problem involves moving object capture and future motion prediction. While object capture specifies which objects are moving in videos, motion prediction describes their future dynamics. Motivated by this analysis, we propose a Cubic Long Short-Term Memory (CubicLSTM) unit for video prediction. CubicLSTM consists of three branches, i.e., a spatial branch for capturing moving objects, a temporal branch for processing motions, and an output branch for combining the first two branches to generate predicted frames. Stacking multiple CubicLSTM units along the spatial branch and output branch, and then evolving along the temporal branch can form a cubic recurrent neural network (CubicRNN). Experiment shows that CubicRNN produces more accurate video predictions than prior methods on both synthetic and real-world datasets.

Hehe Fan, Liang Zheng, Yi Yang

The superiority of deeply learned pedestrian representations has been reported in very recent literature of person re-identification (re-ID). In this paper, we consider the more pragmatic issue of learning a deep feature with no or only a few labels. We propose a progressive unsupervised learning (PUL) method to transfer pretrained deep representations to unseen domains. Our method is easy to implement and can be viewed as an effective baseline for unsupervised re-ID feature learning. Specifically, PUL iterates between 1) pedestrian clustering and 2) fine-tuning of the convolutional neural network (CNN) to improve the original model trained on the irrelevant labeled dataset. Since the clustering results can be very noisy, we add a selection operation between the clustering and fine-tuning. At the beginning when the model is weak, CNN is fine-tuned on a small amount of reliable examples which locate near to cluster centroids in the feature space. As the model becomes stronger in subsequent iterations, more images are being adaptively selected as CNN training samples. Progressively, pedestrian clustering and the CNN model are improved simultaneously until algorithm convergence. This process is naturally formulated as self-paced learning. We then point out promising directions that may lead to further improvement. Extensive experiments on three large-scale re-ID datasets demonstrate that PUL outputs discriminative features that improve the re-ID accuracy.