Junqi Lin

Junqi Lin, Huixin Miao, Junjie Cao et al.

Existing multi-person pose estimators can be roughly divided into two-stage approaches (top-down and bottom-up approaches) and one-stage approaches. The two-stage methods either suffer high computational redundancy for additional person detectors or group keypoints heuristically after predicting all the instance-free keypoints. The recently proposed single-stage methods do not rely on the above two extra stages but have lower performance than the latest bottom-up approaches. In this work, a novel single-stage multi-person pose regression, termed SMPR, is presented. It follows the paradigm of dense prediction and predicts instance-aware keypoints from every location. Besides feature aggregation, we propose better strategies to define positive pose hypotheses for training which all play an important role in dense pose estimation. The network also learns the scores of estimated poses. The pose scoring strategy further improves the pose estimation performance by prioritizing superior poses during non-maximum suppression (NMS). We show that our method not only outperforms existing single-stage methods and but also be competitive with the latest bottom-up methods, with 70.2 AP and 77.5 AP75 on the COCO test-dev pose benchmark. Code is available at https://github.com/cmdi-dlut/SMPR.

Ye Wang, Ziheng Wang, Boshen Xu et al.

Temporal Video Grounding (TVG), the task of locating specific video segments based on language queries, is a core challenge in long-form video understanding. While recent Large Vision-Language Models (LVLMs) have shown early promise in tackling TVG through supervised fine-tuning (SFT), their abilities to generalize remain limited. To address this, we propose a novel post-training framework that enhances the generalization capabilities of LVLMs via reinforcement learning (RL). Specifically, our contributions span three key directions: (1) Time-R1: we introduce a reasoning-guided post-training framework via RL with verifiable reward to enhance the capabilities of LVLMs on the TVG task. (2) TimeRFT: we explore data-efficient post-training strategies on our curated RL-friendly dataset, which trains the model to progressively comprehend difficult samples, leading to better generalization. (3) TVGBench: we carefully construct a small yet comprehensive benchmark for LVLM evaluation, assessing 11 types of queries and featuring balanced distributions across both videos and queries. Extensive experiments demonstrate that Time-R1 achieves state-of-the-art performance across multiple downstream datasets using only 2.5K training data, while improving its general video understanding capabilities.