LOSS RESCALING BY UNCERTAINTY INFERENCE FOR SINGLE-STAGE OBJECT DETECTION

In this paper, we propose a novel loss rescaling method for single-stage convolutional object detection by predicted uncertainty scores. In object detection tasks, the difficulty of each object differs, and most of the background samples are easier to distinguish than others. Equally treating such samples causes data imbalance, resulting in underperforming accuracy and low efficiency in training. To resolve this issue, our method relates prediction difficulty to Bayesian uncertainty estimates and prediction correctness. Subsequently, we modify the architecture of a single-stage object detector to propagate Bayesian uncertainty and use the uncertainty estimates as anchor wise weights to re-scale the training loss. The proposed method is evaluated on both image classification and object detection tasks. The experiment results demonstrate that our method achieves better accuracy than existing loss rescaling methods, and verifies the effectiveness of Bayesian uncertainty inference in improving model training.