Service
Computer vision that measures motion from ordinary phone video
Most real video is filmed on a phone: 30 fps, compressed, moving camera. I build vision systems that measure what this video really contains and say plainly what it does not, instead of returning a confident guess.
Discuss a project →What you get
- Detection, tracking and pose for each person in the frame
- Measurements over time: angles, heights, timing, rotations
- A report where every number carries an uncertainty and a status
- Benchmarks on held-out clips before any change is kept
Where I have already done this
Computer vision · R&DjumpsightA coach films a jump on a phone; the system measures rotations, height and time in the air.±0.54 rotation error on 1,705 jumps (MMFS)91% Axel recognised (FS-Jump3D)3D animation · R&DLodykA coach films an exercise on a phone. The system lifts the movement out of the video and a 3D skater repeats it, so the skating app can teach the exercise without showing the coach's footage or voice.8.50 px arm error, down from 11.800.40% joints bending backwards on unseen clips, down from 2.92%
Questions
- How accurate can it be from one camera?
- In the figure-skating case the rotation error is ±0.54 on 1,705 jumps, and the Axel is recognised 91% of the time. Where the video cannot show something, the system says so.
- Does it need a GPU?
- Not necessarily: one clip is processed in about 13 seconds on a CPU, with the first results after about 2 seconds.