2018/12/06 by Alexandre Ramé, Rame, Alexandre, Emilien Garreau +5
Computer Science · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Face recognition and analysis #Video Surveillance and Tracking Methods
paper · pdf · doi:10.48550/arxiv.1812.02611
openalex publication_date 2018/12/06 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Object detectors tend to perform poorly in new or open domains, and require\nexhaustive yet costly annotations from fully labeled datasets. We aim at\nbenefiting from several datasets with different categories but without\nadditional labelling, not only to increase the number of categories detected,\nbut also to take advantage from transfer learning and to enhance domain\nindependence.\n Our dataset merging procedure starts with training several initial Faster\nR-CNN on the different datasets while considering the complementary datasets'\nimages for domain adaptation. Similarly to self-training methods, the\npredictions of these initial detectors mitigate the missing annotations on the\ncomplementary datasets. The final OMNIA Faster R-CNN is trained with all\ncategories on the union of the datasets enriched by predictions. The joint\ntraining handles unsafe targets with a new classification loss called SoftSig\nin a softly supervised way.\n Experimental results show that in the case of fashion detection for images in\nthe wild, merging Modanet with COCO increases the final performance from 45.5%\nto 57.4% in mAP. Applying our soft distillation to the task of detection with\ndomain shift between GTA and Cityscapes enables to beat the state-of-the-art by\n5.3 points. Our methodology could unlock object detection for real-world\napplications without immense datasets.\n