en.wikipedia.org/w/index.php?title=Activity_recognition&action=history
1 correction found
In computer vision-based activity recognition, fine-grained action localization typically provides per-image segmentation masks delineating the humans, objects and the corresponding action category (e.g., Segment-Tube).
This reverses the usual setup in action localization. Standard spatio-temporal action localization methods typically output frame-level bounding boxes or tubes; Segment-Tube was proposed specifically as a segmentation-based alternative.
Full reasoning
Primary literature describes bounding boxes / action tubes as the standard output for spatio-temporal action localization, not per-frame segmentation masks. The Segment-Tube paper itself says it was inspired by prior localization work using tubelets (sequences of bounding boxes) and proposes segmentation masks instead of bounding boxes. Its introduction also states that spatio-temporal localization is typically done as bounding box regression on 2D frames.
A later ACCV paper likewise summarizes the field by saying that most current spatio-temporal action localization methods use object-detection frameworks and predict frame-level action bounding boxes, which are then linked into tubes.
So the Wikipedia sentence is backwards: segmentation masks are not what this task typically provides; they are a less common, finer-grained variant exemplified by Segment-Tube.
2 sources
- Segment-Tube: Spatio-Temporal Action Localization in Untrimmed Videos with Per-Frame Segmentation
“Inspired by the recent spatio-temporal action localization efforts with tubelets (sequences of bounding boxes), we present ... Segment-tube ... of per-frame segmentation masks.” The paper also says spatio-temporal localization is “typically bounding box regression on 2D frames.”
- Real-time spatio-temporal action localization via learning motion representation
“Most current spatio-temporal action localization methods utilize the popular object detection frameworks, action bounding boxes are predicted in frame-level, then a dynamic linking strategy is used to generate human action tubes.”