Few-Shot Supervised Prototype Alignment for Pedestrian Detection on Fisheye Images

IEEE Conference on Computer Vision and Pattern Recognition(2022)

引用 5|浏览33
Complete and pre-trained models are readily available for download for object detection and can perform well on datasets containing everyday images. Domain adaptation is used to transfer models to more specific datasets with characteristics not present in pre-training. We propose the novel adaptation setting of pedestrian detection in fisheye images, where target samples are scarce but annotated. Our setting provides interesting new challenges for adaptation due to global perspective changes and geometric distortions not found in existing adaptation tasks. To this end, we introduce loss coupling for unsupervised adversarial adaptation and boost prototype-based adaptation with ground-truth information. We additionally propose a novel supervised adaptation head for features in the bounding box regressor. Our method leads to more stable adversarial training and outperforms supervised and unsupervised baselines. Our method requires half the amount of training samples for small datasets to achieve the same performance as supervised fine-tuning.
download,object detection,everyday images,domain adaptation,specific datasets,novel adaptation setting,pedestrian detection,fisheye images,target samples,interesting new challenges,global perspective changes,geometric distortions,adaptation tasks,loss coupling,unsupervised adversarial adaptation,prototype-based adaptation,ground-truth information,supervised adaptation head,stable adversarial training,supervised baselines,unsupervised baselines,training samples,supervised fine-tuning,shot supervised prototype alignment
AI 理解论文
Chat Paper