Navigating the Domain Shift: Object Detection in Indian Road Datasets

Object detection plays a vital role in several applications, including autonomous driving systems and intelligent transportation systems. However, accurate Object detection in unstructured road environments, particularly on Indian roads, poses a significant challenge due to the diverse and complex nature of the surroundings and perpetually changing environment. Moreover, the limited availability of annotated data exacerbates the problem. Real-world deployment of object detection models often introduces a significant challenge: domain shift. To address this issue, we introduced a new augmentation technique called InterAug. The suggested method may use region of an object using a bounding box to define a “semantic context” for each object in contrast to existing policies. InterAug approach illustrates the feasibility of choosing the most suitable context for each object within a specific scene through annotation, as opposed to altering the entire scene or solely specifying bounding boxes. To prove effectiveness of proposed work makes use of an extensive Indian road dataset JUVDsi V1, that includes a variety of terrains, weather patterns, and road kinds. Through extensive experiments and evaluations using benchmark models Faster R-CNN and YOLOV5, we’ve demonstrated our approach’s remarkable effectiveness, achieving a substantial 4.3 \(\%\) accuracy boost compared to the baseline. We address the critical issue of domain shift, emphasizing the importance of day-to-night adaptation in challenging Indian road scenarios.