For 2.1–2.3, training and evaluation use cached ResNet18 image features
(--load_feat). Scores are averaged over 678 test objects.
Each example shows ten matching views around 360°.
Decoder: MLP with two hidden layers of width 512 and ReLU, followed by a linear output layer for 32 × 32 × 32 voxel logits. Loss: binary cross-entropy with logits.
F1@0.05: 40.43
Decoder: MLP with three hidden layers of width 512 and ReLU, followed by a linear output layer for 1,000 XYZ points. Loss: symmetric squared Chamfer distance.
F1@0.05: 78.65
Decoder: MLP with two hidden layers of width 512 and ReLU, predicting vertex offsets for a level-4 sphere with 2,562 vertices. Loss: Chamfer distance plus Laplacian smoothing, with weights 1.0 and 0.1.
F1@0.05: 71.25
Varying n_points across 1,000, 2,500, and 5,000. Each run uses
a two-hidden-layer MLP with width 512 and ReLU, cached ResNet18 features
(--load_feat), and Chamfer loss. Ground-truth surface sampling
remains at 1,000 points. Scores use all predicted points and are averaged
over 678 test objects.
| Predicted points | F1@0.05 |
|---|---|
| 1,000 | 78.51 |
| 2,500 | 83.30 |
| 5,000 | 84.84 |
Conclusion: After seeing how sparse the original 1,000-point predictions were, I expected increasing the point count to recover finer detail. However, the 2,500- and 5,000-point outputs appeared to make the existing coarse shapes denser without clearly recovering additional surface geometry in the visualized examples. Although F1@0.05 increased from 78.51 to 84.84, denser sampling can improve recall by giving more ground-truth samples a predicted point within the matching threshold. This suggests that part of the score increase reflects denser coverage of the coarse shape, rather than a clear improvement in fine reconstruction detail.
The test object is Example 1 from 2.1. Each RGB view is predicted
independently using the same trained voxel MLP with two hidden layers of
width 512 and cached ResNet18 features (--load_feat).
All prediction GIFs use the same ten render angles, camera distance,
elevation, and scale. The ground-truth mesh is shared across all inputs.
Analysis: I wanted to check view consistency by investigating whether the learned decoder predicts similar 3D structure from different views of the same object. Overall, the predictions for this example appear reasonably consistent across the four input views. The seat, backrest, and armrests have a similar arrangement, and the reconstructed chair keeps the same orientation even as the input viewpoint changes.
There are small differences in the width and thickness of the predicted surfaces, but the main structure remains similar. All four predictions miss the chair legs, even though legs are visible in the input images. They also have smoother, more rounded surfaces than the rectangular ground-truth mesh. This shows that view consistency and reconstruction accuracy are separate: the model can produce a stable coarse shape while repeating the same missing details across views. These observations are qualitative and limited to this one test object.
A shared MLP takes an L2-normalized 512-dimensional image feature and an XYZ location and predicts one occupancy logit. The decoder has 3 hidden layers of width 256 with ReLU. All 32 × 32 × 32 locations in the normalized cube (−1, 1)3 are queried for every image during training and evaluation. The ResNet18 image encoder and decoder are trained jointly from RGB using binary cross-entropy with logits.
F1@0.05: 41.24 / 100, averaged over all 678 test objects. F1 compares 1,000 sampled points on each predicted and ground-truth surface. Each example shows ten matching views around 360°.