Average F1@0.05: 37.314
Average F1@0.05: 77.470
Average F1@0.05: 67.735
I varied the mesh Laplacian smoothness weight while keeping the model architecture, learning rate, training steps, and other hyperparameters fixed.
| Mesh smoothness weight (w_smooth) | Average F1 @ 0.05 |
|---|---|
| 0 | 66.491 |
| 0.01 | 66.237 |
| 0.1 | 64.485 |
| 1.0 | 61.104 |
w_smooth=1.0. Stronger smoothing likely
restricts the local vertex deformations needed to represent thin chair
parts and sharper structures such as legs and backrests. The small
weight of 0.01 performed close to the unregularized model, suggesting
weak smoothing provides little accuracy penalty in this experiment.
I reconstructed the same test object from three different input viewpoints to investigate whether the point-cloud model learns a view-invariant 3D representation. Despite substantial changes in the visible surfaces of the chair, the three predicted point clouds are highly similar in their overall geometry and point distribution. This suggests that the network has learned a representation that is relatively insensitive to camera viewpoint rather than simply reproducing the visible 2D silhouette. However, the predictions also appear more similar to one another than the input views might suggest, and they capture only coarse structure rather than the distinctive armrest and back geometry of this particular chair. This indicates that the model may rely strongly on a learned category-level chair prior: it produces a consistent generic 3D structure across views, but some instance-specific geometric details are lost.