🤖 AI Summary
To address degraded detection and localization performance in field robotics vision caused by dynamic shadows, this paper proposes a lightweight, end-to-end differentiable real-time shadow removal method. We introduce the first physics-inspired shadow modeling module explicitly designed for dynamic outdoor scenes, integrated with a dual-branch U-Net variant to enable illumination-invariant feature extraction, adaptive shadow mask prediction, and edge-aware reconstruction. Evaluated on the FieldShadow benchmark, our method achieves 47 FPS on Jetson AGX Orin with a PSNR of 32.7 dB, significantly outperforming existing state-of-the-art approaches. Key contributions include: (i) the first efficient, differentiable shadow modeling framework tailored to dynamic outdoor environments; (ii) a lightweight architecture balancing physical plausibility and millisecond-scale inference; and (iii) substantially improved robustness of visual perception under complex, varying illumination conditions.