SCDet: decoupling discriminative representation for dark object detection via supervised contrastive learning

Despite the significant progress made in object detection algorithms, their potential to operate effectively under the low-light environment remains to be fully explored. Recent methods realize dark object detection on the entire representation of dark images; however, they do not further consider the potential entanglement between dark disturbance and discriminative information in dark images, and thus, the learned representation may be sub-optimal. Towards this issue, we propose supervised contrastive detection (SCDet), a novel unified framework to learn the potential composition of dark images, and decouple the discriminative component for facilitating dark object detection. Specifically, we introduce the dense decoupling contrastive (DDC) pretext task to investigate the feature consistency based on a dark transformation, allowing the learned representation to be independent of the potential entanglement to realize decoupling. Moreover, to further drive the decoupled representation to be discriminative instead of a collapse solution for dark object detection, we incorporate the supervision detection task as an extra optimization objective, resulting in the joint optimization pattern. The two tasks are complementary to each other: the DDC task regularizes the detection to learn more decoupling-friendly representation, while the supervision detection task guides the discriminative representation decoupling. As a result, the SCDet achieves dark object detection by decoding the decoupled discriminative representation of dark images. Extensive experiments on four datasets demonstrate the effectiveness of our method in both synthetic and real-world scenarios. Code is available at https://github.com/TxLin7/SCDet.