[C104] Foundation Model-Guided RGB-to-RAW Generation with Spectral Supervision for RAW-Domain Detection

Abstract

Training RAW-domain object detectors at scale requires large annotated RAW datasets, which are scarce in practice. A promising alternative synthesizes pseudo-RAW images from large-scale RGB datasets via an RGB-to-RAW reconstruction model trained on a small paired set, enabling ISP-free detection without real RAW data. However, existing methods suffer from two limitations. econstruction models trained on small paired datasets overfit to a narrow RGB distribution, degrading when applied to diverse large-scale RGB inputs. Additionally, pixel-space supervision fails to preserve the high-frequency edges and textures that object detectors rely on. We address both limitations by introducing two complementary components into the RGB-to-RAW reconstruction pipeline. First, we replace the jointly trained RGB guidance module with the encoder of a frozen foundation model, providing domain-robust features that generalize across diverse RGB inputs. Second, we introduce a spectral loss that matches 2D FFT magnitudes between reconstructed and target RAW images, explicitly preserving detection-relevant high-frequency content. Extensive experiments on NOD-Nikon, NOD-Sony, and PASCALRAW demonstrate that our method improves detection mAP over existing methods by up to 2.4 mAP when trained solely on synthetic pseudo-RAW without any real RAW supervision, and up to 2.3 mAP when fine-tuned on a small real RAW set across multiple detector architectures.