Dense Reconstruction Visualization

Below we provide an interactive comparison of π³ predicted (pre-trained vs fine-tuned) and ground-truth point maps on five randomly selected scenes from the UnSceneRecon test set. VGGT and WorldMirror results are omitted due to size constraints (but static visualizations are provided in Figure 4 in the accompanying supplementary PDF document). The grey bar shows the scene name and REC in meters of both models (pre-trained and fine-tuned).

As illustrated below, the quality of the dense prediction head is not compromised, even though our alignment scheme does not directly supervise these dense predictions.

Controls:

• Point Map Navigation: Left Click + Drag to Rotate, Right Click + Drag to Pan, Scroll to Zoom.
• Scene Selection: Use Previous/Next buttons or keyboard (A/← for previous, D/→ for next). Jump to a specific scene using the dropdown menu.
• View Images: Click "See Images For Recon" to view the images used for the reconstruction of the current scene.
• Switch View: Click "To RelRot Results" to view pair-wise relative rotation comparisons.