2021/01/21 by Alban Main de Boissiere, de Boissiere, Alban Main, Rita Noumeir +1
Computer Science · #Anomaly Detection Techniques and Applications #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.2101.08851
openalex publication_date 2021/01/21 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Action recognition, early prediction, and online action detection are\ncomplementary disciplines that are often studied independently. Most online\naction detection networks use a pre-trained feature extractor, which might not\nbe optimal for its new task. We address the task-specific feature extraction\nwith a teacher-student framework between the aforementioned disciplines, and a\nnovel training strategy. Our network, Online Knowledge Distillation Action\nDetection network (OKDAD), embeds online early prediction and online temporal\nsegment proposal subnetworks in parallel. Low interclass and high intraclass\nsimilarity are encouraged during teacher training. Knowledge distillation to\nthe OKDAD network is ensured via layer reuse and cosine similarity between\nteacher-student feature vectors. Layer reuse and similarity learning\nsignificantly improve our baseline which uses a generic feature extractor. We\nevaluate our framework on infrared videos from two popular datasets, NTU RGB+D\n(action recognition, early prediction) and PKU MMD (action detection). Unlike\nprevious attempts on those datasets, our student networks perform without any\nknowledge of the future. Even with this added difficulty, we achieve\nstate-of-the-art results on both datasets. Moreover, our networks use infrared\nfrom RGB-D cameras, which we are the first to use for online action detection,\nto our knowledge.\n