Assignment 2

Ziwen Yuan

1

1.1

Optimized Voxel Grid
Ground Truth Voxel Grid

1.2

Optimized Point Cloud
Ground Truth Point Cloud

1.3

Optimized Mesh
Ground Truth Mesh

2

2.1

Example 1: Input RGB
Example 1: Predicted Voxel Grid
Example 1: Ground Truth Mesh
Example 2: Input RGB
Example 2: Predicted Voxel Grid
Example 2: Ground Truth Mesh
Example 3: Input RGB
Example 3: Predicted Voxel Grid
Example 3: Ground Truth Mesh

2.2

Example 1: Input RGB
Example 1: Predicted Point Cloud
Example 1: Ground Truth Mesh
Example 2: Input RGB
Example 2: Predicted Point Cloud
Example 2: Ground Truth Mesh
Example 3: Input RGB
Example 3: Predicted Point Cloud
Example 3: Ground Truth Mesh

2.3

Example 1: Input RGB
Example 1: Predicted Mesh
Example 1: Ground Truth Mesh
Example 2: Input RGB
Example 2: Predicted Mesh
Example 2: Ground Truth Mesh
Example 3: Input RGB
Example 3: Predicted Mesh
Example 3: Ground Truth Mesh

2.4

Voxel Grid F1 Score
Point Cloud F1 Score
Mesh F1 Score
Voxel Grid - Avg F1@0.05: 34.675 Point Cloud - Avg F1@0.05: 81.987 Mesh - Avg F1@0.05: 70.931
Point clouds achieve highest F1 scores due to their ability to capture fine geometric details. Voxel grids suffer from low resolution limitations and blocky representations that poorly approximate curved surfaces. Meshes provide intermediate performance by maintaining surface topology but face challenges with complex geometry reconstruction from single views, and require more training iteration than point clouds. We use same iteration number for all three's training process

2.5

n_points = 2000 vs n_points = 5000 Higher point density (5000 points) significantly improves reconstruction quality by capturing finer geometric details and reducing sparsity artifacts. However, the computational cost increases quadratically with point count, making 2000 points much faster, which can be used for early testing.

2.6

Model Interpretation Example 1
Model Interpretation Example 2
Model Interpretation Example 3
Side-by-side overlay comparison with synchronized rotation: Left (Red) - Predicted point cloud, Right (Blue) - Ground truth point cloud. This visualization enables direct geometric comparison and reveals reconstruction accuracy patterns across different viewpoints.

3

3.3

Extended Dataset F1 Score
Example 1: Input RGB
Example 1: Predicted Point Cloud
Example 1: Ground Truth Mesh
Example 2: Input RGB
Example 2: Predicted Point Cloud
Example 2: Ground Truth Mesh
Example 3: Input RGB
Example 3: Predicted Point Cloud
Example 3: Ground Truth Mesh
Single Class Training (n_points=2000) - Avg F1@0.05: 81.987 Three Class Training (n_points=5000) - Avg F1@0.05: 89.876 Training on extended dataset (chair, car, plane) with higher point density improves average F1 score by 7.889 points. I am not sure whether the improvement is due to increased data diversity across multiple object categories. But it also shows that higher point resolution (5000 vs 2000 points) can result in better performance.

Ziwen Yuan