Separable Auxiliary Training for Real-Time Small Object Detection

During the training of end-to-end detectors, one-to-one label assignments result in an insufficient number of positive samples, impeding the learning of discriminative features. Existing methods have employed one-to-many label assignments and denoising training strategies to provide additional supervision, thereby increasing the number of positive samples or introducing samples with noise. However, these additional supervisions perform bidirectional feature fusion with the original end-to-end models, increasing the computational costs of the model during inference. In this paper, we propose a Separable Auxiliary Training SAT for real-time small object detection to achieve auxiliary supervision without additional inference delay. In SAT, an auxiliary branch supervised by a one-to-many label assignment is adopted to assist a deployment branch during training. To avoid increasing the inference costs, a one-way feature flow from the deployment branch to the auxiliary branch has been designed. The flow ensures that the deployment branch can be deployed independently without sacrificing any accuracy. Extensive experiments demonstrate that SAT can provide additional supervision to enhance performance without increasing computational costs during inference.