Home
Welcome to the DAPs wiki!
We regret to inform that we found an error in the script used to process the action proposals generated by SCNN-proposal network [1]. Temporal action proposals of this method were shifted due to the use of frame indexes belonging to a frame-rate different than the original frame-rate of the video.
We are extremely sorry about this mistake and sincerely apologize for this error. The correct figures of recall analysis appear below. DAPs network does not outperform SCNN-proposal network regarding average recall vs. number of proposals or recall vs. tIoU thresholds for all the points in these curves. It is worth to mention that our approach offers a good trade-off concerning efficiency and efficacy for the task of temporal proposal generation. Our approach provides a computationally efficient way to extract temporal proposals (~2x faster in practice) and achieves competitive average recall compared to SCNN-proposal network. Efficiency is a consequence of not exhaustively exploring multiple temporal scales of action durations.
For those interested in the latest version of these figures, please download the files pointed here.
We appreciate the in-depth analysis of Zhenheng Yang from USC who discovers that using frame indexes provided by [1] yields to miss-aligned temporal proposals.
- Shou, Z., Wang, D., Chang, S. Action temporal localization in untrimmed videos via multi-stage CNNs. IEEE Conference on Computer Vision and Pattern Recognition, CVPR (2016)