[Feature Request - Nok] Localize partially occluded objects using monocular depth estimation #781
Description
✨ Feature description.
Monocular depth estimation can be used to position partially occluded objects, but it requires combining depth prediction with occlusion reasoning, object priors, and uncertainty handling.
💡 Solution description
The system predicts depth for the visible pixels, identifies which pixels belong to the partially occluded object, and then infers the object’s full 3D position using learned shape/size priors.
More concretely:
Monocular depth estimates relative distance for every visible pixel.
Segmentation isolates the visible part of the object.
Occlusion reasoning / amodal completion estimates the hidden extent.
Object priors (typical size and shape) are used to infer the object’s center in 3D.
The final position is taken from the robust average depth of visible parts, corrected by the inferred full object geometry.
🧠 Tip: You can upvote this request by reacting with 👍 to help prioritize it!