VGGT-Ω on three tabletop views
Feed-forward geometry from a single 3-frame forward pass over the left / middle / right
cameras of motherboard_0510_episode_000. One world frame, three recovered
poses, three depth maps — no COLMAP, no bundle adjustment, no depth sensor.
drag to orbit · scroll to zoom · shift-drag to pan
Per-view predictions
Does it actually register?
A point cloud can look convincing while the poses behind it are wrong, because each view is unprojected with its own predicted camera. The test that separates the two: the tabletop is one large plane visible in all three views, so fit a plane to the middle view alone, then measure the other two views against that same plane. If the poses disagree, left and right will float off it.
| View | median |residual| | signed median | within 0.02 |
|---|---|---|---|
| left | 0.0105 | −0.0046 | 86.8% |
| middle | 0.0050 | −0.0039 | 98.6% |
| right | 0.0086 | +0.0071 | 82.2% |
The plane was fit on the middle view only, yet left and right sit on it to a median of about 0.01 — roughly 1% of the scene's extent. What matters more is the signed median: under ±0.007 for every view, so no view is systematically pushed in front of or behind the others. That is the signature of consistent poses. Left and right show a lower “within 0.02” fraction simply because their oblique angle lets them see the hands, the chairs and the floor beyond the table, none of which belong to the plane.
What the model was given, and what it recovered
The three cameras are fixed rig positions observing the same scene at the same instant, so this is a genuine multi-view problem rather than a video. VGGT-Ω was handed only the three RGB frames — the rig's real extrinsics were never supplied, and the RealSense depth stream that exists for this episode was withheld. Both are therefore available as an independent check on the numbers above.
Depth comes out of DenseHead as exp(logits): positive, but fixed
only up to one global scale factor shared across the scene. Distances below are in model
units, not metres, and only their ratios carry meaning. Recovering metric scale
would need one known length in the scene — the motherboard's 244 mm ATX edge
would do it.
Low-confidence pixels are dropped at the percentile shown in the sidebar, and pixels straddling a depth discontinuity are dropped outright: they unproject into thin sheets of invented geometry stretching between a foreground edge and the background behind it. On this scene those edges fall on the hands and the motherboard's outline, which is exactly where they are most visible.