Fig. 2
From: Tubelets: Unsupervised Action Proposals from Spatiotemporal Super-Voxels
Tubelet generation: in the first stage a video is segmented into super-voxels. In addition to segmenting video frames, we also segment their \(iMotion\) maps to also include motion information in the super-voxel segmentation stage. In the second stage of super-voxel grouping, super-voxels are iteratively merged using several grouping functions each of them leading to a set of action proposals. These sets are again grouped by union into a set of Tubelets. The final stage is post-processing that includes pruning and spatiotemporal-refinement of action proposals