Yang Yang

Shicheng Tan, Weng Lam Tam, Yuanchun Wang et al.

Currently, the reduction in the parameter scale of large-scale pre-trained language models (PLMs) through knowledge distillation has greatly facilitated their widespread deployment on various devices. However, the deployment of knowledge distillation systems faces great challenges in real-world industrial-strength applications, which require the use of complex distillation methods on even larger-scale PLMs (over 10B), limited by memory on GPUs and the switching of methods. To overcome these challenges, we propose GKD, a general knowledge distillation framework that supports distillation on larger-scale PLMs using various distillation methods. With GKD, developers can build larger distillation models on memory-limited GPUs and easily switch and combine different distillation methods within a single framework. Experimental results show that GKD can support the distillation of at least 100B-scale PLMs and 25 mainstream methods on 8 NVIDIA A100 (40GB) GPUs.

Yang Yang, Mizuka Komatsu, Kenji Oyama et al.

In stockbreeding of beef cattle, computer vision-based approaches have been widely employed to monitor cattle conditions (e.g. the physical, physiology, and health). To this end, the accurate and effective recognition of cattle action is a prerequisite. Generally, most existing models are confined to individual behavior that uses video-based methods to extract spatial-temporal features for recognizing the individual actions of each cattle. However, there is sociality among cattle and their interaction usually reflects important conditions, e.g. estrus, and also video-based method neglects the real-time capability of the model. Based on this, we tackle the challenging task of real-time recognizing interactions between cattle in a single frame in this paper. The pipeline of our method includes two main modules: Cattle Localization Network and Interaction Recognition Network. At every moment, cattle localization network outputs high-quality interaction proposals from every detected cattle and feeds them into the interaction recognition network with a triple-stream architecture. Such a triple-stream network allows us to fuse different features relevant to recognizing interactions. Specifically, the three kinds of features are a visual feature that extracts the appearance representation of interaction proposals, a geometric feature that reflects the spatial relationship between cattle, and a semantic feature that captures our prior knowledge of the relationship between the individual action and interaction of cattle. In addition, to solve the problem of insufficient quantity of labeled data, we pre-train the model based on self-supervised learning. Qualitative and quantitative evaluation evidences the performance of our framework as an effective method to recognize cattle interaction in real time.

Housheng Xie, Junhui Qiu, Yuan Dai et al.

As an essential processing step before the fusing of infrared and visible images, the performance of image registration determines whether the two images can be fused at correct spatial position. In the actual scenario, the varied imaging devices may lead to a change in perspective or time gap between shots, making significant non-rigid spatial relationship in infrared and visible images. Even if a large number of feature points are matched, the registration accuracy may still be inadequate, affecting the result of image fusion and other vision tasks. To alleviate this problem, we propose a Semantic-Aware on-Demand registration network (SA-DNet), which mainly purpose is to confine the feature matching process to the semantic region of interest (sROI) by designing semantic-aware module (SAM) and HOL-Deep hybrid matching module (HDM). After utilizing TPS to transform infrared and visible images based on the corresponding feature points in sROI, the registered images are fused using image fusion module (IFM) to achieve a fully functional registration and fusion network. Moreover, we point out that for different demands, this type of approach allows us to select semantic objects for feature matching as needed and accomplishes task-specific registration based on specific requirements. To demonstrate the robustness of SA-DNet for non-rigid distortions, we conduct extensive experiments by comparing SA-DNet with five state-of-the-art infrared and visible image feature matching methods, and the experimental results show that our method adapts better to the presence of non-rigid distortions in the images and provides semantically well-registered images.

Yiheng Li, Zichang Tan, Zhen Lei et al.

In AI-generated image detection, current cutting-edge methods typically adapt pre-trained foundation models through partial-parameter fine-tuning. However, these approaches often struggle to generalize to forgeries from unseen generators, as the fine-tuned models capture only limited patterns from training data and fail to reflect the evolving traits of new ones. To overcome this limitation, we propose Image-Adaptive Prompt Learning (IAPL), a novel paradigm that dynamically adjusts the prompts fed into the encoder according to each testing image, rather than fixing them after training. This design significantly enhances robustness and adaptability to diverse forged images. The dynamic prompts integrate conditional information with test-time adaptive tokens through a lightweight learnable scaling factor. The conditional information is produced by a Conditional Information Learner, which leverages CNN-based feature extractors to model both forgery-specific and general conditions. The test-time adaptive tokens are optimized during inference on a single sample by enforcing prediction consistency across multiple views, ensuring that the parameters align with the current image. For the final decision, the optimal input with the highest prediction confidence is selected. Extensive experiments show that IAPL achieves state-of-the-art performance, with mean accuracies of 95.61% and 96.7% on the widely used UniversalFakeDetect and GenImage datasets, respectively. Codes and weights will be released on https://github.com/liyih/IAPL.

Feng Zhang, Zezhong Tan, Xinhong Ma et al.

To combine the advantages of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), recent methods have integrated ''hints'' into post-training, which are prefix segments of complete reasoning trajectories, aiming for powerful knowledge expansion and reasoning generalization. However, existing hint-based RL methods typically ignore difficulty when scheduling hint ratios and estimating relative advantages, leading to unstable learning and excessive imitation of off-policy hints. In this work, we propose ADHint, which treats difficulty as a key factor in both hint-ratio schedule and relative-advantage estimation to achieve a better trade-off between exploration and imitation. Specifically, we propose Adaptive Hint with Sample Difficulty Prior, which evaluates each sample's difficulty under the policy model and accordingly schedules an appropriate hint ratio to guide its rollouts. We also introduce Consistency-based Gradient Modulation and Selective Masking for Hint Preservation to modulate token-level gradients within hints, preventing biased and destructive updates. Additionally, we propose Advantage Estimation with Rollout Difficulty Posterior, which leverages the relative difficulty of rollouts with and without hints to estimate their respective advantages, thereby achieving more balanced updates. Extensive experiments across diverse modalities, model scales, and domains demonstrate that ADHint delivers superior reasoning ability and out-of-distribution generalization, consistently surpassing existing methods in both pass@1 and avg@8. Our code and dataset will be made publicly available upon paper acceptance.

Peng Xu, Shengwu Xiong, Jiajun Zhang et al.

This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year's MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants' methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Hongpeng Pan, Shifeng Yi, Shouwei Yang et al.

This report introduces an enhanced method for the Foundational Few-Shot Object Detection (FSOD) task, leveraging the vision-language model (VLM) for object detection. However, on specific datasets, VLM may encounter the problem where the detected targets are misaligned with the target concepts of interest. This misalignment hinders the zero-shot performance of VLM and the application of fine-tuning methods based on pseudo-labels. To address this issue, we propose the VLM+ framework, which integrates the multimodal large language model (MM-LLM). Specifically, we use MM-LLM to generate a series of referential expressions for each category. Based on the VLM predictions and the given annotations, we select the best referential expression for each category by matching the maximum IoU. Subsequently, we use these referential expressions to generate pseudo-labels for all images in the training set and then combine them with the original labeled data to fine-tune the VLM. Additionally, we employ iterative pseudo-label generation and optimization to further enhance the performance of the VLM. Our approach achieve 32.56 mAP in the final test.

Yang Yang, Jian Wu, Xiangman Song et al.

Hit rate is a key performance metric in predicting process product quality in integrated industrial processes. It represents the percentage of products accepted by downstream processes within a controlled range of quality. However, optimizing hit rate is a non-convex and challenging problem. To address this issue, we propose a data-driven quasi-convex approach that combines factorial hidden Markov models, multitask elastic net, and quasi-convex optimization. Our approach converts the original non-convex problem into a set of convex feasible problems, achieving an optimal hit rate. We verify the convex optimization property and quasi-convex frontier through Monte Carlo simulations and real-world experiments in steel production. Results demonstrate that our approach outperforms classical models, improving hit rates by at least 41.11% and 31.01% on two real datasets. Furthermore, the quasi-convex frontier provides a reference explanation and visualization for the deterioration of solutions obtained by conventional models.

Xinyi Yu, Weiqi He, Xuecheng Qian et al.

Accurate rail location is a crucial part in the railway support driving system for safety monitoring. LiDAR can obtain point clouds that carry 3D information for the railway environment, especially in darkness and terrible weather conditions. In this paper, a real-time rail recognition method based on 3D point clouds is proposed to solve the challenges, such as disorderly, uneven density and large volume of the point clouds. A voxel down-sampling method is first presented for density balanced of railway point clouds, and pyramid partition is designed to divide the 3D scanning area into the voxels with different volumes. Then, a feature encoding module is developed to find the nearest neighbor points and to aggregate their local geometric features for the center point. Finally, a multi-scale neural network is proposed to generate the prediction results of each voxel and the rail location. The experiments are conducted under 9 sequences of 3D point cloud data for the railway. The results show that the method has good performance in detecting straight, curved and other complex topologies rails.

Yang Zhang, Yang Yang, Chenyun Xiong et al.

Encoder-decoder models have been widely used in RGBD semantic segmentation, and most of them are designed via a two-stream network. In general, jointly reasoning the color and geometric information from RGBD is beneficial for semantic segmentation. However, most existing approaches fail to comprehensively utilize multimodal information in both the encoder and decoder. In this paper, we propose a novel attention-based dual supervised decoder for RGBD semantic segmentation. In the encoder, we design a simple yet effective attention-based multimodal fusion module to extract and fuse deeply multi-level paired complementary information. To learn more robust deep representations and rich multi-modal information, we introduce a dual-branch decoder to effectively leverage the correlations and complementary cues of different tasks. Extensive experiments on NYUDv2 and SUN-RGBD datasets demonstrate that our method achieves superior performance against the state-of-the-art methods.

Zongtao Liu, Jing Xu, Jintao Su et al.

Recently, several studies have explored the use of neural network to solve different routing problems, which is an auspicious direction. These studies usually design an encoder-decoder based framework that uses encoder embeddings of nodes and the problem-specific context to produce node sequence(path), and further optimize the produced result on top by beam search. However, existing models can only support node coordinates as input, ignore the self-referential property of the studied routing problems, and lack the consideration about the low reliability in the initial stage of node selection, thus are hard to be applied in real-world. In this paper, we take the orienteering problem as an example to tackle these limitations. We propose a novel combination of a variant beam search algorithm and a learned heuristic for solving the general orienteering problem. We acquire the heuristic with an attention network that takes the distances among nodes as input, and learn it via a reinforcement learning framework. The empirical studies show that our method can surpass a wide range of baselines and achieve results close to the optimal or highly specialized approach. Also, our proposed framework can be easily applied to other routing problems. Our code is publicly available.

Biwei Li, Zhenqiang Mi, Yu Guo et al.

Multi-robot visual simultaneous localization and mapping (SLAM) system is normally consisted of multiple mobile robots equipped with camera and/or other visual sensors. The networked robots work independently or cooperatively in an unknown scene in order to solve autonomous localization and mapping problem. One of the most critical issues in Multi-robot visual SLAM is the intensive computation that is normally required yet overwhelming for inexpensive mobile robots with limited on-board resources. To address this problem, a novel task offloading strategy and dense point cloud map construction method is proposed in this paper. First, we develop a novel strategy to remotely offload computation-intensive tasks to cloud center, so that the tasks that could not originally be achieved locally on the resource-limited robot systems become possible. Second, a modified iterative closest point algorithm (ICP), named fitness score hierarchical ICP algorithm (FS-HICP), is developed to accelerate point cloud registration. The correctness, efficiency, and scalability of the proposed strategy are evaluated with both theoretical analysis and experimental simulations. The results show that the proposed method can effectively reduce the energy consumption while increase the computation capability and speed of the multi-robot visual SLAM system, especially in indoor environment.

Anders Glent Buch, Yang Yang, Norbert Krüger et al.

We present a method for finding correspondence between 3D models. From an initial set of feature correspondences, our method uses a fast voting scheme to separate the inliers from the outliers. The novelty of our method lies in the use of a combination of local and global constraints to determine if a vote should be cast. On a local scale, we use simple, low-level geometric invariants. On a global scale, we apply covariant constraints for finding compatible correspondences. We guide the sampling for collecting voters by downward dependencies on previous voting stages. All of this together results in an accurate matching procedure. We evaluate our algorithm by controlled and comparative testing on different datasets, giving superior performance compared to state of the art methods. In a final experiment, we apply our method for 3D object detection, showing potential use of our method within higher-level vision.