Deep learning for human action recognition in videos is making significantprogress, but is slowed down by its dependency on expensive manual labeling oflarge video collections. In this work, we investigate the generation ofsynthetic training data for action recognition, as it has recently shownpromising results for a variety of other computer vision tasks. We propose aninterpretable parametric generative model of human action videos that relies onprocedural generation and other computer graphics techniques of modern gameengines. We generate a diverse, realistic, and physically plausible dataset ofhuman action videos, called PHAV for "Procedural Human Action Videos". Itcontains a total of 39,982 videos, with more than 1,000 examples for eachaction of 35 categories. Our approach is not limited to existing motion capturesequences, and we procedurally define 14 synthetic actions. We introduce a deepmulti-task representation learning architecture to mix synthetic and realvideos, even if the action categories differ. Our experiments on the UCF101 andHMDB51 benchmarks suggest that combining our large set of synthetic videos withsmall real-world datasets can boost recognition performance, significantlyoutperforming fine-tuning state-of-the-art unsupervised generative models ofvideos.
展开▼
