Enhancing Data Efficiency for Training Object Detectors

Deep learning has transformed object detection in autonomous driving and robotics. Yet, it requires training with large datasets, driving up costs and resource demands. While data reduction techniques offer a solution, most re-search has focused on image classification, leaving object detection largely unexplored. This paper introduces, adapts, and evaluates different data reduction strategies for 2<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D$</tex> object detection. Experiments with Faster R-CNN on nulmages and BDDI00K reveal that: (1) Reduction methods based on loss prove to be both simple and effective, achieving up to 40 % dataset reduction while preserving model performance, (2) We introduce a novel predictive measure for dataset quality, leveraging intrinsic dimensionality to evaluate dataset diversity. This metric achieves up to 80 % alignment (Spearman correlation) with the final performance on the full dataset, enabling efficient pre-evaluation of potential reductions and streamlining the reduction process, (3) We investigate the impact of label errors on data reduction, revealing their influence, especially at high dataset compression rates, and offering key insights for developing robust reduction strategies.