Contrastive Learning-based Robust Object Detection under Smoky Conditions

Abstract

Object detection is to effectively find out interested targets in images and then accurately determine their categories and positions. Recently many excellent methods have been developed to provide powerful detection capability. However, their performance may degrade significantly under severe weather such as smoky conditions. In this paper, we propose a contrastive learning-based robust object detection algorithm for smoke images. The proposed object detector consists of two modules: contrastive learning module and object bounding box prediction module. The first module learns representation vectors by maximizing agreement between different augmented views of the same smoke image. These representations are then sent to the second module to yield the bounding box for each object. In addition, we also propose a novel affine data augmentation method. Extensive experiments have been conducted on A2I2-Haze dataset which is the first real haze dataset with in-situ smoke measurement aligned to aerial and ground imagery. This dataset is also the only dataset used in the 5 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> UG <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> + challenges of CVPR 2022 for both training and testing. Compared with state-of-the-art methods, evaluation results show the superiority of our proposed object detector.