With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning OOD from point annotations has gained great attention. In this paper, we rethink this challenging task setting with the layout among instances and present Point2RBox-v2. At the core are three principles: 1) Gaussian overlap loss. It learns an upper bound for each instance by treating objects as 2D Gaussian distributions and minimizing their overlap. 2) Voronoi watershed loss. It learns a lower bound for each instance through watershed on Voronoi tessellation. 3) Consistency loss. It learns the size/rotation variation between two output sets with respect to an input image and its augmented view. Supplemented by a few devised techniques, e.g. edge loss and copy-paste, the detector is further enhanced. To our best knowledge, Point2RBox-v2 is the first approach to explore the spatial layout among instances for learning point-supervised OOD. Our solution is elegant and lightweight, yet it is expected to give a competitive performance especially in densely packed scenes: 62.61%/86.15%/34.71% on DOTA/HRSC/FAIR1M.

Graph neural networks (GNNs) face scalability challenges. While recent approaches have adopted pooling strategies inspired by convolutional neural networks (CNNs) to reduce graph size and improve efficiency, these methods often focus on local information and are optimized for single graph-level tasks. This limitation hinders their effectiveness in multi-task scenarios that require task-specific global information. We present DOTIN (Dropping Out Task-Irrelevant Nodes), an approach to graph size reductio. DOTIN utilizes K learnable virtual nodes to represent graph embeddings for K distinct graph-level tasks. By employing a transformer-based attention model, it adaptively removes up to 90% of low-attentiveness raw nodes without notable performance degradation. Our method achieves comparable accuracy to state-of-the-art techniques while offering substantial benefits in efficiency. Specifically, DOTIN accelerates Graph Attention Networks (GAT) by approximately 50% on graph-level tasks such as graph classification and graph edit distance (GED). Additionally, it reduces memory usage by about 60% on the D&D dataset. These results show DOTIN's potential to enhance the scalability and efficiency of deep GNNs across multiple graph-level tasks while maintaining high performance.

Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by 6.5%, 7.0%, and 6.7% in three variants of ScanObjectNN with the Mlp-Linear evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE.

Dense visual contrastive learning (DRL) shows promise for learning localized information in dense prediction tasks, but struggles with establishing pixel/patch correspondence across different views (cross-contrasting). Existing methods primarily rely on self-contrasting the same view with variations, limiting input variance and hindering downstream performance. This paper delves into the mechanisms of self-contrasting and cross-contrasting, identifying the crux of the issue: transforming discrete positional embeddings to continuous representations. To address the correspondence problem, we propose a Continuous Relative Rotary Positional Query ({\mname}), enabling patch-level representation learning. Our extensive experiments on standard datasets demonstrate state-of-the-art (SOTA) results. Compared to the previous SOTA method (PQCL), our approach achieves significant improvements on COCO: with 300 epochs of pretraining, {\mname} obtains \textbf{3.4\%} mAP$^{bb}$ and \textbf{2.1\%} mAP$^{mk}$ improvements for detection and segmentation tasks, respectively. Furthermore, {\mname} exhibits faster convergence, achieving \textbf{10.4\%} mAP$^{bb}$ and \textbf{7.9\%} mAP$^{mk}$ improvements over SOTA with just 40 epochs of pretraining.

Reinforcement fine-tuning (RFT) is essential for enhancing the reasoning and generalization capabilities of large language models, but its success heavily relies on the quality of the training data. While data selection has been extensively studied in supervised learning, its role in reinforcement learning, particularly during the RFT stage, remains largely underexplored. In this work, we introduce RFT-Inf, the first influence estimator designed for data in reinforcement learning. RFT-Inf quantifies the importance of each training example by measuring how its removal affects the final training reward, offering a direct estimate of its contribution to model learning. To ensure scalability, we propose a first-order approximation of the RFT-Inf score by backtracking through the optimization process and applying temporal differentiation to the sample-wise influence term, along with a first-order Taylor approximation to adjacent time steps. This yields a lightweight, gradient-based estimator that evaluates the alignment between an individual sample’s gradient and the average gradient direction of all training samples, where a higher degree of alignment implies greater training utility. Extensive experiments demonstrate that RFT-Inf consistently improves reward performance and accelerates convergence in reinforcement fine-tuning.

Long chain-of-thought (LongCoT) has emerged as a powerful reasoning paradigm for enabling large language models (LLMs) to solve complex tasks through a systematic and thorough thinking phase. Although supervised fine-tuning (SFT) on high-quality LongCoT traces has proven effective to activate LongCoT abilities, we find that models trained in this way tend to overfit problem-specific knowledge and heuristics, leading to degraded out-of-distribution performance. To address this issue, we propose a Decoupled LongCoT Fine-Tuning (DLoFT) algorithm, which enables the model to learn generalizable LongCoT reasoning abilities while preventing overfitting to the reasoning content with problem-specific information. The key idea is to decouple the gradient into two orthogonal components: 1) a paradigm-relevant gradient corresponding to the general LongCoT paradigm and 2) a content-relevant gradient reflecting the problem-specific information, where only the former gradient is used to update model parameters. Specifically, by leveraging the unique two-phase composition (thinking and solution) of the LongCoT response, our gradient decoupling mechanism isolates the content-relevant gradient via a projection operation and separates the paradigm-relevant gradient through orthogonalization. Our DLoFT ensures the model concentrate on internalizing the LongCoT paradigm rather than memorizing problem-specific knowledge and heuristics. Extensive experiments demonstrate that our DLoFT significantly improves the generalization behavior of LongCoT abilities compared to SFT while maintaining strong in-distribution performance.

Recent advancements in text-to-video (T2V) diffusion models have enabled high-fidelity and realistic video synthesis. However, current T2V models often struggle to generate physically plausible content due to their limited inherent ability to accurately understand physics. We found that while the representations within T2V models possess some capacity for physics understanding, they lag significantly behind those from recent video self-supervised learning methods. To this end, we propose a novel framework called {VideoREPA}, which distills physics understanding capability from video understanding foundation models into T2V models by aligning token-level relations. This closes the physics understanding gap and enables more physics-plausible generation. Specifically, we introduce the {Token Relation Distillation (TRD) loss}, leveraging spatio-temporal alignment to provide soft guidance suitable for finetuning powerful pre-trained T2V models—a critical departure from prior representation alignment (REPA) methods. To our knowledge, VideoREPA is the first REPA method designed for finetuning T2V models and specifically for injecting physical knowledge. Empirical evaluations show that VideoREPA substantially enhances the physics commonsense of baseline method, CogVideoX, achieving significant improvement on relevant benchmarks and demonstrating a strong capacity for generating videos consistent with intuitive physics. Code and more video results are available at https://videorepa.github.io/.

Image outpainting aims to generate the content of an input sub-image outside its boundaries, which remains open for existing generative models. This paper explores image outpainting in three directions that have not been achieved in literature to our knowledge: outpainting 1) with continuous multiples (in contrast to the discrete ones by existing methods); 2) with arbitrary resolutions; and 3) in a single step (for any multiples and resolutions). The arbitrary multiple outpainting is achieved by utilizing randomly cropped views from the same image during training to capture arbitrary relative positional information. Specifically, by feeding one view and relative positional embeddings as queries, we can reconstruct another view. At inference, we generate images with arbitrary expansion multiples by inputting an anchor image and its corresponding positional embeddings. The continuous-resolution outpainting is achieved by introducing the multi-scale training strategy into generative models. Specifically, by disentangling the image resolution and the number of patches, it can generate images with arbitrary resolutions without post-processing. Meanwhile, we propose a query-based contrastive objective to make our method not rely on a pre-trained backbone network which is otherwise often required in peer methods. The comprehensive experimental results on public benchmarks show its superior performance over state-of-the-art approaches.

Vision transformers (ViTs) have recently been used for visual matching. The original grid dividing strategy of ViTs neglects the spatial information of the keypoints, limiting the sensitivity to local information. We propose QueryTrans (Query Transformer), which adopts a cross-attention module and keypoints-based center crop strategy for better spatial information extraction. We further integrate the graph attention module and devise a transformer-based graph matching approach GMTR (Graph Matching TRansformers) whereby the combinatorial nature of GM is addressed by a graph transformer GM solver. On standard GM benchmarks, GMTR shows competitive performance against the SOTA frameworks. Specifically, on Pascal VOC, GMTR achieves 83.6% accuracy, 0.9% higher than the SOTA. On SPair-71k, GMTR shows great potential and outperforms most of the previous works. Meanwhile, on Pascal VOC, QueryTrans improves the accuracy of NGMv2 from 80.1% to 83.3%, and BBGM from 79.0% to 84.5%. On SPair-71k, it improves NGMv2 from 80.6% to 82.5%, and BBGM from 82.1% to 83.9%. Code is available at: https://github.com/jp-guo/gm-transformer.

Existing image-level self-supervised learning (SSL) methods pre-trained on natural scene data can have difficulty in adating to dense prediction tasks. However, scene images contain multiple varied instances. We devise two techniques to craft high-quality scene and instance views for instance-level SSL. Firstly, we leverage the prior from the image-level pre-trained model to discover the salient instances and the correspondence between cross-image instance pairs. Then, we augment multiple instances to compose the scene views, where both scene similarity and variance are promoted. Meanwhile, we encourage the scene and inner instances to align in feature space to model the scene-instance correlation. Experiments show its SOTA performance in dense prediction tasks.

Image outpainting aims to generate the content of an input sub-image beyond its original boundaries. It is an important task in content generation yet remains an open problem for generative models. This paper pushes the technical frontier of image outpainting in two directions that have not been resolved in literature: 1) outpainting with arbitrary and continuous multiples (without restriction), and 2) outpainting in a single step (even for large expansion multiples). Moreover, we develop a method that does not depend on a pre-trained backbone network, which is in contrast commonly required by the previous SOTA outpainting methods. The arbitrary multiple outpainting is achieved by utilizing randomly cropped views from the same image during training to capture arbitrary relative positional information. Specifically, by feeding one view and positional embeddings as queries, we can reconstruct another view. At inference, we generate images with arbitrary expansion multiples by inputting an anchor image and its corresponding positional embeddings. The one-step outpainting ability here is particularly noteworthy in contrast to previous methods that need to be performed for $N$ times to obtain a final multiple which is $N$ times of its basic and fixed multiple. We evaluate the proposed approach (called PQDiff as we adopt a diffusion-based generator as our embodiment, under our proposed \textbf{P}ositional \textbf{Q}uery scheme) on public benchmarks, demonstrating its superior performance over state-of-the-art approaches. Specifically, PQDiff achieves state-of-the-art FID scores on the Scenery (\textbf{21.512}), Building Facades (\textbf{25.310}), and WikiArts (\textbf{36.212}) datasets. Furthermore, under the 2.25x, 5x and 11.7x outpainting settings, PQDiff only takes \textbf{40.6\%}, \textbf{20.3\%} and \textbf{10.2\%} of the time of the benchmark state-of-the-art (SOTA) method.

We propose a novel text-to-video (T2V) generation benchmark, ChronoMagic-Bench, to evaluate the temporal and metamorphic knowledge skills in time-lapse video generation of the T2V models (e.g. Sora and Lumiere). Compared to existing benchmarks that focus on visual quality and text relevance of generated videos, ChronoMagic-Bench focuses on the models’ ability to generate time-lapse videos with significant metamorphic amplitude and temporal coherence. The benchmark probes T2V models for their physics, biology, and chemistry capabilities, in a free-form text control. For these purposes, ChronoMagic-Bench introduces 1,649 prompts and real-world videos as references, categorized into four major types of time-lapse videos: biological, human creation, meteorological, and physical phenomena, which are further divided into 75 subcategories. This categorization ensures a comprehensive evaluation of the models’ capacity to handle diverse and complex transformations. To accurately align human preference on the benchmark, we introduce two new automatic metrics, MTScore and CHScore, to evaluate the videos' metamorphic attributes and temporal coherence. MTScore measures the metamorphic amplitude, reflecting the degree of change over time, while CHScore assesses the temporal coherence, ensuring the generated videos maintain logical progression and continuity. Based on the ChronoMagic-Bench, we conduct comprehensive manual evaluations of eighteen representative T2V models, revealing their strengths and weaknesses across different categories of prompts, providing a thorough evaluation framework that addresses current gaps in video generation research. More encouragingly, we create a large-scale ChronoMagic-Pro dataset, containing 460k high-quality pairs of 720p time-lapse videos and detailed captions. Each caption ensures high physical content and large metamorphic amplitude, which have a far-reaching impact on the video generation community. The source data and code are publicly available on https://pku-yuangroup.github.io/ChronoMagic-Bench.

Masked autoencoder has been widely explored in point cloud self-supervised learning, whereby the point cloud is generally divided into visible and masked parts. These methods typically include an encoder accepting visible patches (normalized) and corresponding patch centers (position) as input, with the decoder accepting the output of the encoder and the centers (position) of the masked parts to reconstruct each point in the masked patches. Then, the pre-trained encoders are used for downstream tasks. In this paper, we show a motivating empirical result that when directly feeding the centers of masked patches to the decoder without information from the encoder, it still reconstructs well. In other words, the centers of patches are important and the reconstruction objective does not necessarily rely on representations of the encoder, thus preventing the encoder from learning semantic representations. Based on this key observation, we propose a simple yet effective method, $i.e.$, learning to \textbf{P}redict \textbf{C}enters for \textbf{P}oint \textbf{M}asked \textbf{A}uto\textbf{E}ncoders (\textbf{PCP-MAE}) which guides the model to learn to predict the significant centers and use the predicted centers to replace the directly provided centers. Specifically, we propose a Predicting Center Module (PCM) that shares parameters with the original encoder with extra cross-attention to predict centers. Our method is of high pre-training efficiency compared to other alternatives and achieves great improvement over Point-MAE, particularly surpassing it by \textbf{5.50\% on OBJ-BG, 6.03\% on OBJ-ONLY, and 5.17\% on PB-T50-RS} for 3D object classification on the ScanObjectNN dataset. The code is available at \url{https://github.com/aHapBean/PCP-MAE}.

Dense contrastive learning (DCL) has been recently explored for learning localized information for dense prediction tasks (e.g., detection and segmentation). It still suffers the difficulty of mining pixels/patches correspondence between two views. A simple way is inputting the same view twice and aligning the pixel/patch representation. However, it would reduce the variance of inputs, and hurts the performance. We propose a plug-in method PQCL (Positional Query for patch-level Contrastive Learning), which allows performing patch-level contrasts between two views with exact patch correspondence. Besides, by using positional queries, PQCL increases the variance of inputs, to enhance training. We apply PQCL to popular transformer-based CL frameworks (DINO and iBOT, and evaluate them on classification, detection and segmentation tasks, where our method obtains stable improvements, especially for dense tasks. It achieves new state-of-the-art in most settings. Code is available at https://github.com/Sherrylone/Query_Contrastive.

Model pre-training is essential in human-centric perception. In this paper, we first introduce masked image modeling (MIM) as a pre-training approach for this task. Upon revisiting the MIM training strategy, we reveal that human structure priors offer significant potential. Motivated by this insight, we further incorporate an intuitive human structure prior - human parts - into pre-training. Specifically, we employ this prior to guide the mask sampling process. Image patches, corresponding to human part regions, have high priority to be masked out. This encourages the model to concentrate more on body structure information during pre-training, yielding substantial benefits across a range of human-centric perception tasks. To further capture human characteristics, we propose a structure-invariant alignment loss that enforces different masked views, guided by the human part prior, to be closely aligned for the same image. We term the entire method as HAP. HAP simply uses a plain ViT as the encoder yet establishes new state-of-the-art performance on 11 human-centric benchmarks, and on-par result on one dataset. For example, HAP achieves 78.1% mAP on MSMT17 for person re-identification, 86.54% mA on PA-100K for pedestrian attribute recognition, 78.2% AP on MS COCO for 2D pose estimation, and 56.0 PA-MPJPE on 3DPW for 3D pose and shape estimation.

Existing symmetric contrastive learning methods suffer from collapses (complete and dimensional) or quadratic complexity of objectives. Departure from these methods which maximize mutual information of two generated views, along either instance or feature dimension, the proposed paradigm introduces intermediate variables at the feature level, and maximizes the consistency between variables and representations of each view. Specifically, the proposed intermediate variables are the nearest group of base vectors to representations. Hence, we call the proposed method ARB (Align Representations with Base). Compared with other symmetric approaches, ARB 1) does not require negative pairs, which leads the complexity of the overall objective function is in linear order, 2) reduces feature redundancy, increasing the information density of training samples, 3) is more robust to output dimension size, which out-performs previous feature-wise arts over 28% Top-1 accuracy on ImageNet-100under low-dimension settings.

Negative pairs, especially hard negatives as combined with common negatives (easy to discriminate), are essential in contrastive learning, which plays a role of avoiding degenerate solutions in the sense of constant representation across different instances. Inspired by recent hard negative mining methods via pairwise mixup operation in vision, we propose M-Mix, which dynamically generates a sequence of hard negatives. Compared with previous methods, M-Mix mainly has three features: 1) adaptively choose samples to mix; 2) simultaneously mix multiple samples; 3) automatically assign different mixing weights to the selected samples. We evaluate our method on two image datasets (CIFAR-10, CIFAR-100), five node classification datasets (PPI, DBLP, Pubmed, etc), five graph classification datasets (IMDB, PTC_MR, etc), and two downstream combinatorial tasks (graph edit distance and node clustering). Results show that it achieves state-of-the-art performance under self-supervised settings. Code is available at: https://github.com/Sherrylone/m-mix.

Adversarial attacks are very successful on image classification, but there are few researches on vision-language systems, such as image captioning. In this paper, we study the robustness of a CNN+RNN based image captioning system being subjected to adversarial noises in complex domain. In particular, we propose Fooled-by-Imagination, a novel algorithm for crafting adversarial examples with semantic embedding of targeted caption as perturbation in complex domain. The proposed algorithm explores the great merit of complex values in introducing imaginary part for modeling adversarial perturbation, and maintains the similarity of the image in real part. Our approach provides two evaluation approaches, which check whether neural image captioning systems can be fooled to output some randomly chosen captions or keywords. Besides, our method has good transferability under black-box setting. At last, our extensive experiments show that our algorithm can successfully craft visually-similar adversarial examples with randomly targeted captions or keywords at a higher success rate.

Ensemble is a general way of improving the accuracy and stability of learning models, especially for the generalization ability on small datasets. Compared with tree-based methods, relatively less works have been devoted to an in-depth study on effective ensemble design for neural networks. In this paper, we propose a principled ensemble technique by constructing the so-called diversified ensemble layer to combine multiple networks as individual modules. We theoretically show that each individual model in our ensemble layer corresponds to weights in the ensemble layer optimized in different directions. Meanwhile, the devised ensemble layer can be readily integrated into popular neural architectures, including CNNs, RNNs, and GCNs. Extensive experiments are conducted on public tabular datasets, images, and texts. By adopting weight sharing approach, the results show our method can notably improve the accuracy and stability of the original neural networks with ignorable extra time and space overhead.