Hand Mesh Recovery from Images
Hand recovery fits an articulated hand model to high-resolution crops because fingers need detail that body crops erase.
Why Does This Exist?
Body crops give hands about pixels, so fingers blur into mittens. Hands also articulate degrees of freedom with self-occlusion. Dedicated hand recovery crops at pixels or more and fits MANO (shape , pose for joints). SMPL-X attaches the result to the body.
Think of It Like This
Reading rings from a jewel close-up, not a street photo
You cannot read ring engravings from across the street. You walk up, take a macro shot, and trace each finger separately. The street photo finds the person; the macro reads the hand.
Body crops are street photos and hand crops are macros. The analogy stops at interaction: hands touch objects while rings sit still.
How It Actually Works
A detector crops the hand at or higher. The net regresses MANO plus weak camera. Twenty-one joints project to D with loss under pixels on visible tips. Contact loss penalizes mesh interpenetration past mm for grasping hands.
Worked example
An open palm crop shows all tips. Index tip projects within pixels, pinky within , wrist within . Thumb flexion rad matches the photo angle. Closing into a fist hides four tips; visible wrist plus silhouette keep error at pixels while hidden joints rely on prior with -degree uncertainty. That gap is the honest limit.
Watch Out For
Motion blur merging adjacent fingers
A s exposure during a wave smears index into middle finger. The symptom is webbed meshes on video frames. Raise shutter speed or drop blurred frames from fitting.
The Quick Version
- Hands need dedicated high-resolution crops, not body-crop leftovers.
- MANO compresses articulation into pose plus shape like SMPL does.
- Twenty-one joints anchor fingers within a few pixels when visible.
- Do not trust hidden finger angles; priors guess with wide uncertainty.
- Contact terms stop grasping fingers passing through each other.