Manipulation action recognition is a challenging problem in computer vision. We previously reported a system based on matching groups of superpixels. In this paper, we modify the superpixel group mining algorithm and report results on two datasets. Recognition accuracies are comparable with those reported using deep learning. The representation used in our approach is amenable to interpretation. Specifically, visualisation of matched groups provides a level of explanation for recognition decisions and insights into the likely generalisation ability of action representations.
|Title of host publication||ReaLX 2018|
|Subtitle of host publication||Proceedings of the SICSA Workshop on Reasoning, Learning and Explainability|
|Editors||Kyle Martin , Nirmalie Wiratunga, Leslie S. Smith|
|Number of pages||5|
|Publication status||Published - 27 Jul 2018|
|Name||CEUR Workshop Proceedings|
- Action recognition
- Computer vision
- Superpixel group mining
Huang, T., & McKenna, S. (2018). Superpixel Group Mining for Manipulation Action Recognition. In K. Martin , N. Wiratunga, & L. S. Smith (Eds.), ReaLX 2018: Proceedings of the SICSA Workshop on Reasoning, Learning and Explainability (Vol. 2151, pp. 1-5). (CEUR Workshop Proceedings). CEUR-WS.