FluxMark
Recovering forensic watermarks from surfaces that refuse to stay still — clothing, jerseys, fabric in motion — by unwrapping the deforming surface into a stable UV domain before attempting to read the mark.
The problem
Watermark recovery assumes a surface that behaves. Embed a mark in a flat plane and you can find it again, because the geometry between embedding and reading is roughly the same.
A jersey on a moving body breaks that assumption completely. The surface stretches, compresses, folds, self-occludes and rotates out of view, all within a few frames. A mark painted onto that surface is not degraded so much as resampled by an unknown, non-rigid transform every frame. Read it in image space and there is nothing stable to correlate against.
The approach
Stop reading in image space. DensePose gives a per-pixel correspondence from the body in the image to a canonical UV parameterisation of a human surface. Invert it and the deforming garment can be resampled into that canonical domain — where the mark sits still, flat and rectangular, frame after frame, regardless of what the body is doing.
Recovery then becomes a tractable problem on a stable image, and accumulating across frames adds signal instead of smearing it. The same mapping runs in reverse: a surface you can read a mark from is a surface you can also write to, which turns a jersey in a live broadcast into addressable space for marking or in-scene placement.
Where it stands
There is a working implementation. The unwrap runs in real time and mark recovery from non-rigid surfaces is demonstrated end to end. The manuscript is in preparation, so this has not been peer reviewed.
Related
The Media Intelligent Engine handles the other half: choosing where in a frame a mark can safely live, and embedding it during packaging.