Optimizing Object Detection via Metric-Driven Training Data Selection

Abstract

In the realm of object detection, training models with limited, unlabelled data from target domains presents significant challenges. This study focuses on the critical issue of optimizing image dataset selection to enhance object detection performance, especially when dealing with small sample sizes and closely-related target data that lacks predefined labels. Our proposed method adopts an integrated approach that combines data exploration, pseudo labeling, and strategic image selection from varied datasets, e.g. COCO and KITTI. By ranking source images based on their image-wise Average Precision (AP) scores followed by mosaic augmentation on selected images, experimental results demonstrate the efficiency of this data selection mechanism, indicating significant advancements in object detection performance and domain adaptability. Our method won the 2nd DataCV Challenge with the AP of 0.2285, achieving a 0.052 AP increase over the baseline method. This work offers a robust pathway to overcome key challenges in applying object detection models across various domains, particularly in scenarios with limited annotations from target set. Our codes have been available at: https://github.com/welovecv/datacv.