Interactive demo

Values are read verbatim from the metrics.txt written by eval_two_view.py --two-view-test for this particular pair. The qualitative mode only selects a pair whose rotation and translation errors both fall below 3 degrees, so these are per-pair figures, not dataset-level accuracies. Dataset-level RRA@15 and RTA@15 for every benchmark are in Table 1 of the repository README.

Both clouds are the fused two-view overlay: view 0 in its own camera frame, plus view 1 mapped into that frame by the predicted relative pose. They are raw network outputs with no global alignment or post-processing, subsampled to 120k points for the web. CAM3R recovers geometry up to a global scale, so each cloud is recentred on its bounding sphere and normalised to unit radius before display; that puts prediction and ground truth in a common frame and makes the shape comparison the meaningful one. The .ply files written by the evaluation script keep their original coordinates.

Each cloud is subsampled to 100k points for the web and is drawn in the coordinate frame its own method writes out, with no rotation applied, so the four panels can start at different orientations. The methods reconstruct at very different scales, so every cloud is recentred on its bounding sphere and scaled to unit radius at render time only; the .ply files keep their original coordinates.