Mean Teacher for Unsupervised Domain Adaptation in Multi-View 3D Pedestrian Detection
Resumo
We introduce an innovative unsupervised domain adaptation method designed to enhance the performance of multi-view 3D pedestrian detection in unlabeled target scenes. Our approach revolves around tailoring the Mean Teacher architecture to suit multi-view scenarios. A distinctive aspect of our work involves the proposal of a novel multi-view data augmentation procedure specifically crafted for integration into our Mean Teacher framework. Additionally, we introduce an optional step that takes into account the field of view of augmented outputs during the computation of our unsupervised loss. To assess the efficacy of our proposed solution, we conducted evaluations on the widely-used WILDTRACK and MultiviewX datasets. The results demonstrated superior performance compared to state-of-the-art unlabeled techniques. Notably, our method achieved an accuracy that surpassed the best unlabeled method by 4.2 percentage points and 0.3 percentage points when applied to WILDTRACK and MultiviewX as target datasets, respectively. These findings underscore the effectiveness of our unsupervised domain adaptation method in enhancing multi-view 3D pedestrian detection, showcasing its potential for real-world applications.
