Case Study 1 / Factory Egocentric Tag Annotation
One frame. Three levels. Every action accounted for.
Egocentric video annotation for worker activity recognition and process optimization.
- Capture
- Head-mounted fisheye
- Video
- 20+ hours
- Scale
- ~1M frames
- Platform
- CVAT
Turning first-person factory footage into structured activity data.
The project uses head-mounted fisheye video recorded from a worker's point of view in a factory. Frames are tagged at three levels: the person's overall activity, the right-hand action, and the left-hand action. The work is performed frame by frame, with careful attention to short actions and transitions.
| CAPTURE Head-mounted fisheye | VIDEO VOLUME 20+ hours | FRAME SCALE ~1M frames | PLATFORM CVAT |
| PERSON LEVEL Walk · Assembly · Load · Unload · Inspect | RIGHT HAND Transport · Use · Rest · Hold | LEFT HAND Transport · Use · Rest · Hold | OUTPUT Tagged video frames |
From raw footage to a validated frame.
The three-phase visual uses the same factory footage to show the difference between an untouched frame, an annotation state that needs correction, and a valid final frame.
Original fisheye factory frame.
Original video frame before tag labels are applied.
Three tag levels. One tag from each.
| 01 / PERSON Walk · Assembly · Load · Unload · Inspect Describes the worker's overall activity in the frame. | 02 / RIGHT HAND Transport · Use · Rest · Hold Captures the action performed by the right hand. | 03 / LEFT HAND Transport · Use · Rest · Hold Captures the action performed by the left hand. |
| VALID FRAME 3 tags total — 1 Person + 1 Right Hand + 1 Left Hand | REVIEW REQUIRED 2 or 4+ tags, a missing level, or multiple tags from the same level |
Short actions must be captured precisely. A hand may move through transport, hold, and use over only a few frames, so tag transitions need to follow the action visible in each frame.
Not just tags. A consistent frame.
Each frame must contain exactly three tags: one Person-level tag, one Right-hand tag, and one Left-hand tag. Frames with missing tags, extra tags, or duplicate tags within a level are flagged for correction.
Quality control checks tag count, tag-level uniqueness, and whether each tag correctly describes the action in the frame.
Frame-level labels for worker activity recognition.
The project produces structured, frame-by-frame tag annotations intended to support worker activity recognition and process optimization. The workflow is managed in CVAT.
| ONE FRAME Every action accounted for | THREE LEVELS Person · Right hand · Left hand | ONE CAREFUL PROCESS Frame by frame, in CVAT |
