Unreliability-aware Disentangling for Cross-Domain Semi-supervised Pedestrian Detection

Wenhao Wu, Si Wu, Hau-San Wong; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2510-2526


The rapid progress of pedestrian detection is supported by the ever-growing labeled training data and elaborate neural-network-based model. However, adequate labeled training data are not always accessible when it comes to a new scene. Semi-supervised learning is promising for the case where a small amount of manually annotated images and a large amount of unannotated images are handy. In the semi-supervised setting, data generation is a powerful technique as a type of data augmentation. Some methods conduct data generation by disentangling pedestrian instances into different codes in latent space and combining codes of different instances to reconstruct new instances. However, these methods either work in a single domain or cannot handle the case where some instances are partially represented in the images. In this work, we propose to solve code-level information transferring from reliable domains to unreliable domains by incorporating a domain classifier that competes with the disentangling module to generate domain-invariant codes. An external classifier is trained on appearance-enhanced instances and sends integrity signals to the generative module, which facilitates the generative module to recognize fully/partially represented pedestrian instances. The resulting classifier ultimately renders high-quality pseudo-annotations for the unannotated data. The pseudo-annotated data, combined with a small amount of manually annotated data, are used to achieve a detector with more generalization and accuracy. We perform extensive experiments on multiple challenging benchmarks to demonstrate the effectiveness of the proposed method.

Related Material

@InProceedings{Wu_2022_ACCV, author = {Wu, Wenhao and Wu, Si and Wong, Hau-San}, title = {Unreliability-aware Disentangling for Cross-Domain Semi-supervised Pedestrian Detection}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {2510-2526} }