Name: Joshua Tsai
Andrew ID: joshuat3
Due to limited computational resources, all experiments were trained for 10,000 iterations for consistency and fair comparison.
We optimize a randomly initialized voxel grid using binary cross-entropy loss. The optimized voxel occupancy is compared with the ground-truth voxel grid.
Optimized voxel grid and ground-truth voxel grid.
We optimize a randomly initialized point cloud using Chamfer distance. The optimized point cloud captures the geometry of the target object.
Optimized point cloud and ground-truth point cloud.
We optimize the source mesh using Chamfer loss on sampled surface points together with Laplacian smoothness regularization. The optimized mesh captures the overall geometry of the target chair while maintaining a relatively smooth surface.
Optimized mesh and ground-truth mesh.
We use a pretrained ResNet-18 encoder to extract a 512-dimensional image feature, followed by a 3D transposed-convolution decoder to predict a \(32 \times 32 \times 32\) voxel occupancy grid. The voxel model achieves an average F1@0.05 of 57.033 on the test set. The predicted voxel grids generally capture the overall shape and volume of the chairs, including the seat, backrest, and armrests. However, thin structures such as chair legs and fine geometric details are often missing or overly thick. This is mainly due to the limited spatial resolution of the \(32^3\) voxel representation, which introduces discretization and quantization errors.
Qualitative voxel reconstruction results. The model captures the overall chair shape, while thin structures and fine details are often lost due to the limited voxel resolution.
F1 score under different distance thresholds. The voxel model achieves F1@0.05 = 57.033.
A pretrained ResNet-18 is used as the image encoder, followed by an MLP decoder that directly predicts 1,000 3D points. The model is trained with Chamfer distance for 10,000 iterations. The final average F1@0.05 is 79.393.
Qualitative point cloud reconstruction results. The model captures the overall chair geometry well, although some predicted points are scattered outside the object surface.
F1 score under different distance thresholds. The point cloud model achieves F1@0.05 = 79.393.
A pretrained ResNet-18 is used as the image encoder. Starting from an icosphere, an MLP decoder predicts an offset for each mesh vertex. The model is trained with Chamfer distance and mesh smoothness loss for 10,000 iterations. The final average F1@0.05 is 72.715.
Qualitative mesh reconstruction results. The model recovers the overall chair shape with a continuous surface, but some regions are distorted due to the fixed mesh topology.
F1 score under different distance thresholds. The mesh model achieves F1@0.05 = 72.715.
We vary the mesh smoothness weight \(w_{\text{smooth}}\) while keeping the other training settings unchanged.
| w_smooth | F1@0.05 |
|---|---|
| 0.1 | 72.715 |
| 0.5 | 71.047 |
Increasing the smoothness weight produces smoother and more regular mesh surfaces, but slightly reduces reconstruction accuracy. A larger \(w_{\text{smooth}}\) restricts vertex deformation, making it harder for the mesh to fit detailed chair geometry.
Comparison of mesh reconstructions with different smoothness weights. Compared with the \(w_{\text{smooth}}=0.1\) results shown in Q2.3, increasing the weight to 0.5 reduces irregular deformations and produces smoother surfaces, but may oversmooth geometric details.
F1 score under different distance thresholds for the mesh model with \(w_{\text{smooth}}=0.5\). The model achieves F1@0.05 = 71.047, slightly lower than the 72.715 obtained with \(w_{\text{smooth}}=0.1\) in Q2.3.
To better understand the learned mesh model, I compare the three highest-F1 and three lowest-F1 test examples. The high-F1 cases are mostly chairs with relatively simple and regular geometry, where the model successfully captures the overall seat, backrest, and leg structure. In contrast, the low-F1 cases often contain duplicated objects, unusual viewpoints, thin structures, or more complex geometry. These cases lead to severe mesh distortion and incorrect surface deformation. This suggests that the model learns the coarse shape of common chair structures well, but the fixed-topology mesh deformation has difficulty handling unusual geometry, multiple objects, and thin structures.
Three high-F1 reconstruction examples (98.646, 98.546, and 98.196). The model performs well on chairs with relatively regular geometry and clearly visible major structures.
Three low-F1 reconstruction examples (12.316, 13.828, and 15.261). Failure cases include unusual geometry, multiple objects, and thin structures, which lead to severe mesh distortion.