Seeing in RAW: Detection-Oriented Representation Learning for Low-Light Object Detection

Low-light object detection remains challenging because conventional image signal processing (ISP) pipelines may suppress weak sensor-level cues before detection. Existing enhancement methods usually operate on ISP-processed RGB images, making it difficult to recover subtle textures, intensity variations, and inter-channel responses degraded by denoising, tone mapping, and compression. To address this issue, this paper proposes a native four-channel Bayer RAW-domain framework that learns detection-oriented representations directly from RAW sensor data. Local RAW Signal Modulation (LRSM) enhances weak intensity and structural responses in dark regions, while Detection-Oriented RAW Signal Adaptation (DRSA) adapts multi-scale RAW features under detection-loss supervision. Experiments on RAW datasets captured with Canon, Nikon, and Sony cameras show consistent improvements over existing methods under diverse low-light conditions.