2020/10/27 by R. Gnana Praveen, R, Gnana Praveen, Éric Granger +3
Computer Science · Psychology · #Computer Vision and Pattern Recognition (cs.CV) #Emotion and Mood Recognition #FOS: Computer and information sciences #Face recognition and analysis #Human Pose and Action Recognition #Sleep and Work-Related Fatigue
paper · pdf · doi:10.48550/arxiv.2010.15675
openalex publication_date 2020/10/27 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Automatic estimation of pain intensity from facial expressions in videos has\nan immense potential in health care applications. However, domain adaptation\n(DA) is needed to alleviate the problem of domain shifts that typically occurs\nbetween video data captured in source and target do-mains. Given the laborious\ntask of collecting and annotating videos, and the subjective bias due to\nambiguity among adjacent intensity levels, weakly-supervised learning (WSL)is\ngaining attention in such applications. Yet, most state-of-the-art WSL models\nare typically formulated as regression problems, and do not leverage the\nordinal relation between intensity levels, nor the temporal coherence of\nmultiple consecutive frames. This paper introduces a new deep learn-ing model\nfor weakly-supervised DA with ordinal regression(WSDA-OR), where videos in\ntarget domain have coarse la-bels provided on a periodic basis. The WSDA-OR\nmodel enforces ordinal relationships among the intensity levels as-signed to\nthe target sequences, and associates multiple relevant frames to sequence-level\nlabels (instead of a single frame). In particular, it learns discriminant and\ndomain-invariant feature representations by integrating multiple in-stance\nlearning with deep adversarial DA, where soft Gaussian labels are used to\nefficiently represent the weak ordinal sequence-level labels from the target\ndomain. The proposed approach was validated on the RECOLA video dataset as\nfully-labeled source domain, and UNBC-McMaster video data as weakly-labeled\ntarget domain. We have also validated WSDA-OR on BIOVID and Fatigue (private)\ndatasets for sequence level estimation. Experimental results indicate that our\napproach can provide a significant improvement over the state-of-the-art\nmodels, allowing to achieve a greater localization accuracy.\n