Raw or Cooked? Object Detection on RAW Images

Images fed to a deep neural network have in general undergone several handcrafted image signal processing (ISP) operations, all of which been optimized produce visually pleasing images. In this work, we investigate the hypothesis that intermediate representation images is sub-optimal for downstream computer vision tasks compared RAW representation. We suggest operations ISP instead should be towards end task, by learning parameters jointly during training. extend previous works on topic and propose new learnable operation enables an object detector achieve superior performance when both traditional RGB experiments open PASCALRAW dataset, empirically confirm our hypothesis.