This research tool is for evaluating temporal segmentation algorithms.
Temporal segmentation is a process of cutting sequential data into segments with different semantic meanings. Evaluation of temporal segmentation is done by comparing segments generated by an algorithm of interest with segments in a ground truth.
Download (Official Version): Microsoft Store (to be updated)
Requirements: Windows 10+
Download (Portable Version): ver1.0 (0.8 MB)
Requirements: Windows (any desktop)
Discussions: discussions on Little Series
Privacy Policy: read
As final performance metrics, this tool have Perf and F1r. Perf [1] is a combination of F1-Score and Accurate Temporal Segmentation Rate (ATSR). It considers not only two typically used rates, Precision (P) and Recall (R) constituting the F1-Score, but also a frame-level precision in locating key frames.
F1r is similar to Perf, except that concordance rate (r) [2] is used instead of ATSR. This rate ensure that the output values between 0 and 1, while ATSR does not. In addition, ATSR suffers when the difference in sizes of segment between GT and Alg is large
Methods used are based on the paper below:
[1] ChairGest: a challenge for multimodal mid-air gesture recognition for close HCI. In Proceedings of the 15th ACM on International conference on multimodal interaction 2013 Dec 9 (pp. 483-488). ACM.
[2] Li M, Leung H. Graph-based representation learning for automatic human motion segmentation. Multimedia Tools and Applications. 2016:1-20.
Evaluation Result
