Q1¶

Q1.1¶

Visualize the optimized voxel grid along-side the ground truth voxel grid using the tools learnt in previous section. (Left GT, Right Predicted)

Chair GT Voxel Chair predicted Voxel

Q1.2¶

Visualize the optimized point cloud along-side the ground truth point cloud using the tools learnt in previous section. (Left GT, Right Predicted)

Chair GT PC Chair predicted PC

Q1.3¶

Visualize the optimized mesh along-side the ground truth mesh using the tools learnt in previous section. (Left GT, Right Predicted)

Chair GT Mesh Chair predicted Mesh

Q2¶

Q2.1¶

Note: Trained on GPU, no load_feat used.

[3 EXAMPLES: Image | Ground Truth | Predicted]

Chair GT image Chair GT Vox Chair predicted Vox

Chair GT image Chair GT Vox Chair predicted Vox

Chair GT image Chair GT Vox Chair predicted Vox

Q2.2¶

Note: Trained on GPU, no load_feat used.

[3 EXAMPLES: Image | Ground Truth | Predicted]

Chair GT image Chair GT Point Chair predicted Point

Chair GT image Chair GT Point Chair predicted Point

Chair GT image Chair GT Point Chair predicted Point

Q2.3¶

Note: Trained on GPU, no load_feat used.

[3 EXAMPLES: Image | Ground Truth | Predicted]

Chair GT image Chair GT mesh Chair predicted mesh

Chair GT image Chair GT mesh Chair predicted mesh

Chair GT image Chair GT mesh Chair predicted mesh

Q2.4¶

Quantitatively, the F1 score for the point cloud yields better results than the voxel grid or mesh representation across different thresholds. After point clouds, the model performs better on the mesh representation, and the voxel grid is last. This makes sense because point clouds are the simplest 3D representation, so there’s less ambiguity about where each 3D point should go - it’s mostly just regressing points to the correct locations. Meshes have two objectives (Chamfer distance and smoothness), which help produce a better reconstruction. The voxel grid, on the other hand, is the least expressive, resulting in the worst reconstruction among the three.

Eval vox Eval Point CEval Mesh

Q2.5¶

Change: n_points = 200

When I reduced n_points from 1000 to 200, the F1 score decreased, but the predicted points were more deliberate and meaningful. With 1000 points, many predictions clustered near the chair’s center especially where the back legs meet the seat. This created a dense pooling effect. In contrast, using fewer points forced the model to distribute points more evenly across the object, reducing that central clustering. However, because the model had fewer points to represent the full shape, overall coverage declined, leading to a lower F1 score.

Eval point Chair Image Overlay

Q2.6¶

These visualizations pair input images with predicted and ground-truth 3D point clouds, providing an intuitive view of how well the model captures object geometry beyond numerical metrics:

Chair Image Chair PC GT Overaly

Chair Image Chair PC GT Overaly

Chair Image Chair PC GT Overaly

Note: These were made with the same model used in Q2.2

Q3¶

Q3.1¶

In this problem I implemented a implict decoder that takes 3D locations as inputs along with an image and predicts the occupancy of said 3d locations.

Question 3.1