.png){kind=icon}
.png){kind=icon}
Return
Action temporal detection method based on confidence curve analysis
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-020-08771-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Action temporal detection is a derivative task of action recognition which needs researchers to predict temporal intervals and specific categories in untrimmed videos. Aiming at the problem of too many proposed segments and insufficient filtering effect in multi-stage networks, we propose an action temporal detection method using confidence curve analysis to generate proposal segments. Fixed step window sliding is adopted to generate candidate segments in a video, and we adjust a training mode in segment network. The proposal segments are generated by analyzing the confidence curve of candidate segments, finally proposal segments are input into localization network to classify and adjust confidence level. Extensive experiments performed on THUMOS2014 benchmark show that the proposed method performs significantly better than the original muti-stage convolutional network that mAP increase from 19.0% to 26.4% with 252% accelerating.
Keywords:
Temporal detection
Proposed segment
Confidence curve
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W