Frame Augmentation for Imbalanced Object Detection Datasets
Nada Hesham Kamaledin Elasal, David M. Swart, Nicholas Miller · Journal of Computational Vision and Imaging Systems · 2018
A major challenge in most object detection datasets is class imbal-ance. It is especially apparent in uncurated datasets where framesoriginate from a real-world setup such as a set of cameras col-lecting data from fixed locations. In that case, the dataset classdistribution mirrors the real-world distribution, causing a bias to-wards over-represented classes if used for model training. In thispaper we propose a synthesis technique for balancing the dataset,which exploits having sets of frames from the same camera view.The result is synthesized frames containing only rare objects, whileguaranteeing realistic object placement both in terms of scene con-text and perspective. We train a deep learning object detectionmodel on the augmented dataset and compare its performance toa model trained on the original, imbalanced dataset. Results showthat including the synthesized frames in the training results in asignificant performance boost for the rare classes.