TY - JOUR AU - Elasal, Nada AU - Swart, David M. AU - Miller, Nicholas PY - 2018/12/24 Y2 - 2024/03/29 TI - Frame Augmentation for Imbalanced Object Detection Datasets JF - Journal of Computational Vision and Imaging Systems JA - J. Comp. Vis. Imag. Sys. VL - 4 IS - 1 SE - Articles DO - UR - https://openjournals.uwaterloo.ca/index.php/vsl/article/view/341 SP - 3 AB - <p>A major challenge in most object detection datasets is class imbal-<br>ance. It is especially apparent in uncurated datasets where frames<br>originate from a real-world setup such as a set of cameras col-<br>lecting data from fixed locations. In that case, the dataset class<br>distribution mirrors the real-world distribution, causing a bias to-<br>wards over-represented classes if used for model training. In this<br>paper we propose a synthesis technique for balancing the dataset,<br>which exploits having sets of frames from the same camera view.<br>The result is synthesized frames containing only rare objects, while<br>guaranteeing realistic object placement both in terms of scene con-<br>text and perspective. We train a deep learning object detection<br>model on the augmented dataset and compare its performance to<br>a model trained on the original, imbalanced dataset. Results show<br>that including the synthesized frames in the training results in a<br>significant performance boost for the rare classes.</p> ER -