Scholarlink

<p>Robust object detection remains challenging under adverse conditions due to occlusion, scale changes, and bad weather (such as rain and fog). Although state-of-the-art object detection (such as YOLOv8m) achieves a strong baseline performance, it suffers from a significant performance drop under these challenging conditions because it lacks explicit strategies for addressing occluded regions and scale-dependent feature variances. Therefore, this research proposes a lightweight enhancement of YOLOv8m by considering two lightweight attention modules: the Learnable Occlusion-Aware Module (LOAM) and the Depth Point-wise Multi-Scale Channel Attention (DPMS). LOAM utilizes learnable spatial masks to automatically suppress the occluded region, and DPMS refines channel importance by exploring multi-scale context to address scale variation and noise disturbance. The proposed method is intended to enhance the representation abilities of the feature map and adds very little computational cost. Experiments show that the proposed method increases F1 score by 3.85% (from 75.09% to 78.94%) on the DAWN dataset and recall by 3.71% (from 67.49% to 71.02%) in the complex real-world driving environment of 580 Freeway in San Francisco Bay. Notably, this performance gain comes with minimal computational overhead, increasing the average inference time by only 2.49 ms–from 40.02 ms to 42.51 ms on an NVIDIA GeForce RTX 3050 GPU.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLO-LOAMDPMS: Learnable Occlusion-Aware and Depth Point-wise Multi-Scale Channel Attention for Object Detection

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Nhat Minh Nguyen,
  • Vinh Dinh Nguyen

摘要

Robust object detection remains challenging under adverse conditions due to occlusion, scale changes, and bad weather (such as rain and fog). Although state-of-the-art object detection (such as YOLOv8m) achieves a strong baseline performance, it suffers from a significant performance drop under these challenging conditions because it lacks explicit strategies for addressing occluded regions and scale-dependent feature variances. Therefore, this research proposes a lightweight enhancement of YOLOv8m by considering two lightweight attention modules: the Learnable Occlusion-Aware Module (LOAM) and the Depth Point-wise Multi-Scale Channel Attention (DPMS). LOAM utilizes learnable spatial masks to automatically suppress the occluded region, and DPMS refines channel importance by exploring multi-scale context to address scale variation and noise disturbance. The proposed method is intended to enhance the representation abilities of the feature map and adds very little computational cost. Experiments show that the proposed method increases F1 score by 3.85% (from 75.09% to 78.94%) on the DAWN dataset and recall by 3.71% (from 67.49% to 71.02%) in the complex real-world driving environment of 580 Freeway in San Francisco Bay. Notably, this performance gain comes with minimal computational overhead, increasing the average inference time by only 2.49 ms–from 40.02 ms to 42.51 ms on an NVIDIA GeForce RTX 3050 GPU.