In this work, we introduce an intelligent video sensor for the problem of Action Proposals (AP). AP consists of localizing temporal segments in untrimmed videos that are likely to contain actions. Solving this problem can accelerate several video action understanding tasks, such as detection, retrieval, or indexing. All previous AP approaches are supervised and offline, i.e. they need both the temporal annotations of the datasets during training and access to the whole video to effectively cast the proposals. We propose here a new approach which, unlike the rest of the state-of-the-art models, is unsupervised. This implies that we do not allow it to see any labeled data during learning nor to work with any pre-trained feature on the used dataset. Moreover, our approach also operates in an online manner, which can be beneficial for many real-world applications where the video has to be processed as soon as it arrives at the sensor, e.g., robotics or video monitoring. The core of our method is based on a Support Vector Classifier (SVC) module which produces candidate segments for AP by distinguishing between sets of contiguous video frames. We further propose a mechanism to refine and filter those candidate segments. This filter optimizes a learning-to-rank formulation over the dynamics of the segments. An extensive experimental evaluation is conducted on Thumos'14 and ActivityNet datasets, and, to the best of our knowledge, this work supposes the first unsupervised approach on these main AP benchmarks. Finally, we also provide a thorough comparison to the current state-of-the-art supervised AP approaches. We achieve 41% and 59% of the performance of the best-supervised model on ActivityNet and Thumos'14, respectively, confirming our unsupervised solution as a correct option to tackle the AP problem. The code to reproduce all our results will be publicly released upon acceptance of the paper.
Keywords: action proposals; action recognition; computer vision; intelligent video sensor; unsupervised learning.