Regularizing deep networks with label geometry for accurate object localization on small training datasets

Localization is a critical subtask in object detection, which is closely related to spatial information of objects. Most current detectors simply rely on the fitting ability of deep neural networks to regress towards numerical targets such as coordinates of object boxes. Training deep networks for sufficient fitting ability requires a large number of annotations that are expensive to obtain. In this work, we fully exploit limited annotations by extracting label geometry to improve localization performance on small datasets. We generate distance transform of bounding box edges according to localization labels, with which we supervise intermediate outputs of networks pixel by pixel to reconstruct object geometry for localization. Distance transform is sensitive to box edges and provides geometric gradients flowing into boundaries. We learn such gradients to enhance geometric-aware features through a coupled training with regression, and use it to refine regressed boxes in an evolutionary manner in inference. Extensive experiments are implemented to demonstrate the effectiveness of our method. Our method can be applied in applications that required human-machine interaction, such as the driver-assistance system in autonomous driving, by providing accurate detections to assist humans in making better decisions.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Localization is a critical subtask in object detection, and this work proposes a method to improve localization performance on small datasets by fully exploiting limited annotations. By extracting label geometry and generating distance transform, the method reconstructs object geometry through pixel-wise supervision, and enhances geometric-aware features through coupled training with regression.