Scholarlink

<p>Stereo matching, which mimics human binocular vision for depth perception, remains challenging when balancing real-time performance and high accuracy on both high-end GPUs and resource-constrained edge devices. To address this, we present FASNet—a high-precision, real-time two-stage stereo matching network featuring several key innovations. First, we introduce an Enhanced Residual Block with L2 normalization (ERB-L2), which replaces standard convolutions with depthwise separable convolutions and lightweight L2 normalization, enhancing feature extraction while reducing complexity. Second, we propose a 3D Residual Aggregation (3DRA) module and a Gate-Attention 2D Aggregation (GA2DA) module that combine 3D residual blocks with lightweight gate-attention based 2D fusion to improve detail capture and disparity estimation, while mitigating the cost of traditional 3D convolutions. Third, we design an Entropy-guided Disparity Refinement (EDR) strategy, which uses prediction entropy as an uncertainty metric to linearly refine initial disparities, boosting accuracy with minimal overhead. FASNet processes 1232 × 368 stereo pairs at 243 FPS on an NVIDIA 4090 GPU, and achieves real-time 62 FPS on Jetson AGX Orin with TensorRT. Experiments show that FASNet delivers state-of-the-art speed and accuracy, making it ideal for real-time applications.</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

FASNet: a high-precision, two-stage stereo matching network for real-time depth perception

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Bifa Liang,
  • Yichao Wang,
  • Ziyang Hu,
  • Zhicong Huang,
  • Haifeng Hu,
  • Jianming Xu,
  • Dihu Chen

摘要

Stereo matching, which mimics human binocular vision for depth perception, remains challenging when balancing real-time performance and high accuracy on both high-end GPUs and resource-constrained edge devices. To address this, we present FASNet—a high-precision, real-time two-stage stereo matching network featuring several key innovations. First, we introduce an Enhanced Residual Block with L2 normalization (ERB-L2), which replaces standard convolutions with depthwise separable convolutions and lightweight L2 normalization, enhancing feature extraction while reducing complexity. Second, we propose a 3D Residual Aggregation (3DRA) module and a Gate-Attention 2D Aggregation (GA2DA) module that combine 3D residual blocks with lightweight gate-attention based 2D fusion to improve detail capture and disparity estimation, while mitigating the cost of traditional 3D convolutions. Third, we design an Entropy-guided Disparity Refinement (EDR) strategy, which uses prediction entropy as an uncertainty metric to linearly refine initial disparities, boosting accuracy with minimal overhead. FASNet processes 1232 × 368 stereo pairs at 243 FPS on an NVIDIA 4090 GPU, and achieves real-time 62 FPS on Jetson AGX Orin with TensorRT. Experiments show that FASNet delivers state-of-the-art speed and accuracy, making it ideal for real-time applications.