Motion-Aware Reasoning from Speech to Mask Tracks: Runner-up Solution for the MeViS-Audio Track of the 8th LSVOS Challenge 2026
本文提出Speech2MaskTrack方法，通过语音识别、运动中心时间定位、掩模跟踪等步骤，解决语音引导的视频对象分割问题。
0 citationsRead paper
Selected work
Representative Papers
High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching
Existing single-step generative visuomotor policies often suffer from spatial bias, loss of high-frequency details, and ambiguous action modes due to oversimplified inference. To address these limitations, this work proposes a high-fidelity single-step framework that integrates Recursive Consistent Action Flow (RCAF) to correct spatial truncation errors, Dual-Timestep Frequency Consistency (DTFC) to preserve high-frequency manipulation details, and Contrastive Flow Matching (CFM) to disentangle multimodal action distributions. Requiring only a single forward pass (1 NFE), the method achieves or surpasses the control performance of multi-step baselines—such as 10-step policies—on RoboTwin, Adroit, DexArt, and real robotic platforms, significantly reducing latency while enhancing both precision and action diversity.
0 citationsRead paper
SNAT-YOLO: Efficient Cross-Layer Aggregation Network for Edge-Oriented Gangue Detection
To address the challenges of slow inference speed, low accuracy, and poor deployability of deep learning models for coal-gangue detection on industrial edge devices, this paper proposes a lightweight and efficient object detection framework. Methodologically: (i) we design ADown, a lightweight downsampling module; (ii) we construct C2PSA-TriAtt, a cross-layer feature aggregation architecture integrating Triplet Attention; and (iii) we introduce Inner-FocalerIoU, a novel loss function that jointly optimizes localization accuracy and convergence on hard samples. Built upon YOLOv11, the framework adopts ShuffleNetV2 as its backbone and incorporates all three innovations. Experiments demonstrate a state-of-the-art mAP of 99.10%, with a 38% reduction in model size, 41% fewer parameters, 40% lower computational cost (FLOPs), and a 1 ms decrease in per-image inference latency—significantly fulfilling real-time deployment requirements on coal-mine edge devices.
0 citationsRead paper
Recent publications
Latest Papers
Motion-Aware Reasoning from Speech to Mask Tracks: Runner-up Solution for the MeViS-Audio Track of the 8th LSVOS Challenge 2026
本文提出Speech2MaskTrack方法，通过语音识别、运动中心时间定位、掩模跟踪等步骤，解决语音引导的视频对象分割问题。
0 citationsRead paper
High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching
Existing single-step generative visuomotor policies often suffer from spatial bias, loss of high-frequency details, and ambiguous action modes due to oversimplified inference. To address these limitations, this work proposes a high-fidelity single-step framework that integrates Recursive Consistent Action Flow (RCAF) to correct spatial truncation errors, Dual-Timestep Frequency Consistency (DTFC) to preserve high-frequency manipulation details, and Contrastive Flow Matching (CFM) to disentangle multimodal action distributions. Requiring only a single forward pass (1 NFE), the method achieves or surpasses the control performance of multi-step baselines—such as 10-step policies—on RoboTwin, Adroit, DexArt, and real robotic platforms, significantly reducing latency while enhancing both precision and action diversity.
0 citationsRead paper
SNAT-YOLO: Efficient Cross-Layer Aggregation Network for Edge-Oriented Gangue Detection
To address the challenges of slow inference speed, low accuracy, and poor deployability of deep learning models for coal-gangue detection on industrial edge devices, this paper proposes a lightweight and efficient object detection framework. Methodologically: (i) we design ADown, a lightweight downsampling module; (ii) we construct C2PSA-TriAtt, a cross-layer feature aggregation architecture integrating Triplet Attention; and (iii) we introduce Inner-FocalerIoU, a novel loss function that jointly optimizes localization accuracy and convergence on hard samples. Built upon YOLOv11, the framework adopts ShuffleNetV2 as its backbone and incorporates all three innovations. Experiments demonstrate a state-of-the-art mAP of 99.10%, with a 38% reduction in model size, 41% fewer parameters, 40% lower computational cost (FLOPs), and a 1 ms decrease in per-image inference latency—significantly fulfilling real-time deployment requirements on coal-mine edge devices.
0 citationsRead paper