We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:


References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computer Vision and Pattern Recognition

Title: Budget-Aware Activity Detection with A Recurrent Policy Network

Abstract: In this paper, we address the challenging problem of efficient temporal activity detection in untrimmed long videos. While most recent work has focused and advanced the detection accuracy, the inference time can take seconds to minutes in processing each single video, which is too slow to be useful in real-world settings. This motivates the proposed budget-aware framework, which learns to perform activity detection by intelligently selecting a small subset of frames according to a specified time budget. We formulate this problem as a Markov decision process, and adopt a recurrent network to model the frame selection policy. We derive a recurrent policy gradient based approach to approximate the gradient of the non-decomposable and non-differentiable objective defined in our problem. In the extensive experiments, we achieve competitive detection accuracy, and more importantly, our approach is able to substantially reduce computation time and detect multiple activities with only 0.35s for each untrimmed long video.
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:1712.00097 [cs.CV]
  (or arXiv:1712.00097v2 [cs.CV] for this version)

Submission history

From: Xiaodong Yang [view email]
[v1] Thu, 30 Nov 2017 21:52:27 GMT (2856kb,D)
[v2] Tue, 8 May 2018 04:55:57 GMT (3180kb,D)

Link back to: arXiv, form interface, contact.