Learning Codebook Weights for Action Detection

Abstract

In this work we present a discriminative codebook weighting approach for action detection. We learn global and local weights for the codewords by considering the spatio-temporal Hough voting space of the training sequences. In contrast to the Implicit Shape Model (ISM) where all the codewords that are matched to a local descriptor cast votes with uniform weights, we learn local weights for the matched codewords. In order to learn the local weights we employ Locality-constrained Linear Coding (LLC). Further, we formulate the learning of the global weights as a convex quadratic programming and use alternating optimization to solve for the weights. We demonstrate the performance of the algorithm on KTH action dataset where we compare with the Hough detector using kmeans codebook.

Cite

Text

Kumar and Patras. "Learning Codebook Weights for Action Detection." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2012. doi:10.1109/CVPRW.2012.6239257

Markdown

[Kumar and Patras. "Learning Codebook Weights for Action Detection." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2012.](https://mlanthology.org/cvprw/2012/kumar2012cvprw-learning/) doi:10.1109/CVPRW.2012.6239257

BibTeX

@inproceedings{kumar2012cvprw-learning,
  title     = {{Learning Codebook Weights for Action Detection}},
  author    = {Kumar, B. G. Vijay and Patras, Ioannis},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  year      = {2012},
  pages     = {27-32},
  doi       = {10.1109/CVPRW.2012.6239257},
  url       = {https://mlanthology.org/cvprw/2012/kumar2012cvprw-learning/}
}