Dynamic R-CNN: Towards High Quality Object Detection via Dynamic Training

TL;DR

Dynamic R-CNN introduces a training method that adaptively adjusts label assignment and regression loss parameters during training, significantly improving high-quality object detection performance without extra computational cost.

Contribution

It proposes a novel dynamic training approach that automatically adjusts training criteria based on proposal statistics, addressing the inconsistency problem in fixed settings.

Findings

Achieves 1.9% AP improvement over baseline

Improves AP90 by 5.5% on MS COCO

No extra computational overhead

Abstract

Although two-stage object detectors have continuously advanced the state-of-the-art performance in recent years, the training process itself is far from crystal. In this work, we first point out the inconsistency problem between the fixed network settings and the dynamic training procedure, which greatly affects the performance. For example, the fixed label assignment strategy and regression loss function cannot fit the distribution change of proposals and thus are harmful to training high quality detectors. Consequently, we propose Dynamic R-CNN to adjust the label assignment criteria (IoU threshold) and the shape of regression loss function (parameters of SmoothL1 Loss) automatically based on the statistics of proposals during training. This dynamic design makes better use of the training samples and pushes the detector to fit more high quality samples. Specifically, our method improves upon ResNet-50-FPN baseline with 1.9% AP and 5.5% AP$_{90}$ on the MS COCO dataset with no extra overhead. Codes and models are available at https://github.com/hkzhang95/DynamicRCNN.