Towards Balanced Learning for Instance Recognition

Instance recognition is rapidly advanced along with the developments of various deep convolutional neural networks. Compared to architectures networks, training process, which also crucial success detectors, has received relatively less attention. In this work, we carefully revisit standard practice and find that detection performance often limited by imbalance during generally consists in three levels - sample level, feature objective level. To mitigate adverse effects caused thereby, propose Libra R-CNN, a simple yet effective framework towards balanced learning for instance recognition. It integrates IoU-balanced sampling, pyramid, re-weighting, respectively reducing at sample, feature, Extensive experiments conducted on MS COCO, LVIS Pascal VOC datasets prove effectiveness overall design.