[16-889] Learning for 3D Vision, Spring 2022

Assignment 2

• Number of late days used - 4

1. Practicing with Cameras

1.1. Fitting a voxel grid (5 points)

1.2. Fitting a point cloud (10 points)

1.3. Fitting a mesh (5 points)

2. Reconstructing 3D from single view

2.1. Image to voxel grid (15 points)

Example 1

Example 2

Example 3

2.2. Image to point cloud (15 points)

Example 1

Example 2

Example 3

2.3. Image to mesh (15 points)

Example 1

Example 2

Example 3

2.4. Quantitative comparisions(10 points)

Voxel: Avg F1@0.05: 3.213

Mesh: Avg F1@0.05: 89.949

Pointcloud: Avg F1@0.05: 89.928

2.5. Analyse effects of hyperparms variations (10 points)

Changing w_chamfer

With increasing chamfer loss, the shape of the object seems to be more well defined while the one with the lowest chamfer loss weight is very noisy (corresponding connectivity of vertices are also not smooth).

w_chamfer = 1

w_chamfer = 10

w_chamfer = 100

2.6. Interpret your model (15 points)

Computing the normals of point sets is a good way to visualize the kind of surfaces the network tries to learn when predicting pointclouds. The normals were estimated based on a 10-neighborhood radius and plotted.