Image Decomposition for Object Detection in Low-light Environment

Object detection is one of the most popular and explored vision-based applications. Most object detection approaches focus on the classification and localization of objects in images with proper visibility. However, the visibility of the image may sometimes degrade due to the unavailability of proper lightness (low-light), haze in surroundings and many other environmental conditions. This paper focuses on analyzing the performance of object detection methods in low-light conditions. RetinaNet, FCOS and generalized focal loss are considered to analyse the impact of low-light in object detection out of many object detection methods. In this work, SA-OD, and DSE-OD (decomposition-based methods) are developed for object detection in low-light conditions. SA-OD method estimates a structure-aware reflectance followed by the detection method. DSE-OD estimates the reflectance using the deep network followed by the object detection method. These methods achieve better detection performance in low-light conditions. The low-light images are generated using multiple gamma values and then used for object detection to analyse the performance.