In this project, we propose a novel approach for image augmentation and object detection using the Rain CycleGAN and YOLO (You Only Look Once) algorithms. The goal is to improve the performance of object detection models under adverse weather conditions, specifically rain. We utilize the BDD100k dataset, which contains a large collection of urban driving scenes captured in various weather conditions, including rain.
First, we employ the Rain CycleGAN, a specialized variant of the CycleGAN architecture, to generate synthetic rainy images from the existing BDD100k dataset. The Rain CycleGAN is trained to transform clear weather images into realistic rainy scenes, thereby augmenting the dataset with a diverse range of weather conditions. This augmentation aims to enhance the robustness and generalization of subsequent object detection models, enabling them to effectively handle rainy environments.
Next, we integrate the augmented dataset into the YOLO framework, a popular real-time object detection algorithm known for its efficiency and accuracy. YOLO operates by dividing an image into a grid and predicting bounding boxes and class probabilities for each grid cell. By training YOLO on the augmented dataset, we aim to improve its ability to detect and classify objects accurately in rainy conditions.
Through this combined approach, we expect to achieve significant advancements in object detection performance under challenging weather conditions. The results obtained from our experiments will provide valuable insights into the effectiveness of the Rain CycleGAN and YOLO algorithms for augmenting and detecting objects in rainy urban driving scenes.