Motion-Conditioned Identifiability and Abstention for Physical 3D Scene Geometry
When can camera motion settle what a scene physically is? A real corridor, its mirror reflection and a photo of it on a screen can produce one image; this work picks the motion that separates them, or abstains.
One image, three worlds
Scroll to watch it unfold. Each beat is one step of the question the paper asks.
- 01
One image
A real corridor, its mirror reflection and a photo of it on a screen can all produce the same picture.
- 02
Rival explanations
Each physical explanation of the scene is kept as a candidate instead of being collapsed into one guess.
- 03
Choose the motion
Among the camera motions available, it picks the one that best separates the closest rival explanations.
- 04
Settled
Real depth, a mirror and a flat screen move differently under that motion, so the scene’s explanation can be identified.
- 05
Or abstain
When no available motion can separate the rivals, it abstains rather than commit to an answer it cannot support.
A schematic of the idea, drawn from the write-up below. It shows no results from the paper.
Overview
Single images are ambiguous about the physical world behind them. A real corridor, its reflection in a mirror and a photo of it shown on a screen can produce one and the same image, and a model that reads 3D geometry from that image has no way to tell which world it is looking at.
This work asks a sharper question than “what is the depth here?”: can moving the camera settle the scene’s physical explanation at all, and if so, which motion does it?
Approach
The rival explanations of a scene are kept side by side. The system then chooses the camera motion that best separates the closest of them, because real depth, a reflection and a flat display respond to motion in different ways.
Identification is treated as something to earn, not assume: when no available motion can tell the rivals apart, the system abstains instead of reporting a confident geometry that the evidence does not support.
Status
Under double-blind review. The paper, code and results will be posted here once the review period ends.