Scholarlink

<p>High-resolution fisheye surround-view systems (SVS) generate data streams that severely strain in-vehicle networks. Conventional codecs optimize pixel-level distortion (e.g., PSNR), fundamentally misaligning with machine perception under strict bitrate constraints. Furthermore, prevailing perception pipelines rely on Bird’s-Eye-View (BEV) projection, introducing interpolation artifacts and peripheral field-of-view loss in distorted fisheye imagery. This study presents SVS-JPC, a rectification-free, task-aware framework for joint compression and multi-camera fusion. By optimizing a rate–distortion–perception objective directly in the latent space, intermediate pixel-domain reconstruction is eliminated. Structured feature interaction is achieved via a cross-view latent attention module, bypassing explicit geometric rectification. Evaluated on the SynWoodscape benchmark at 0.128 bits per pixel (bpp), SVS-JPC achieves a 13 percentage-point mean Average Precision (mAP) improvement over HEVC tandem pipelines, reaching 58.29% compared to a 57.30% uncompressed baseline. A reproducible perception improvement over the uncompressed baseline emerges in this low-bitrate regime, attributed to task-aware entropy regularization that suppresses task-irrelevant high-frequency signals. Under zero-shot sim-to-real evaluation on WoodScape, consistent gains over codec-driven baselines demonstrate enhanced robustness to domain shift.</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

Rectification-Free Task-Aware Joint Compression for Fisheye Surround-View Systems via Cross-View Latent Attention

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Basem Barakat,
  • Muhammad A. M. Islam,
  • Mohammed E. M. Rasmy,
  • ELSayed Hemayed

摘要

High-resolution fisheye surround-view systems (SVS) generate data streams that severely strain in-vehicle networks. Conventional codecs optimize pixel-level distortion (e.g., PSNR), fundamentally misaligning with machine perception under strict bitrate constraints. Furthermore, prevailing perception pipelines rely on Bird’s-Eye-View (BEV) projection, introducing interpolation artifacts and peripheral field-of-view loss in distorted fisheye imagery. This study presents SVS-JPC, a rectification-free, task-aware framework for joint compression and multi-camera fusion. By optimizing a rate–distortion–perception objective directly in the latent space, intermediate pixel-domain reconstruction is eliminated. Structured feature interaction is achieved via a cross-view latent attention module, bypassing explicit geometric rectification. Evaluated on the SynWoodscape benchmark at 0.128 bits per pixel (bpp), SVS-JPC achieves a 13 percentage-point mean Average Precision (mAP) improvement over HEVC tandem pipelines, reaching 58.29% compared to a 57.30% uncompressed baseline. A reproducible perception improvement over the uncompressed baseline emerges in this low-bitrate regime, attributed to task-aware entropy regularization that suppresses task-irrelevant high-frequency signals. Under zero-shot sim-to-real evaluation on WoodScape, consistent gains over codec-driven baselines demonstrate enhanced robustness to domain shift.