Home Work 1: FNU Abhimanyu (abhiman2@andrew.cmu.edu)

Collaborated with Ananya Bal

1.1 Voxel fitting

The gif on the top is the generated voxel and the source voxel

1.2 Point cloud fitting

The gif on the top is the generated point cloud and the source point cloud

1.3 Mesh cloud fitting

The gif on the top is the generated mesh and the source mesh

2.1 Voxel fitting

2.4 F1 scores for Mesh|Point|Vox @ 0.05

For mesh the F1 score at 0.05 is:86.coulc For point the F1 score at 0.05 is: 88.27 For vox the F1 score at 0.05 is: 79.45 The mesh has a lower F1 score than the point cloud, because of lack in recreating the connectivity of the original mesh. Point cloud has a lower score because it doesn't need to model these connectivities. The voxel has the lowest F1 score, because it couldn't recreate thin components like the leg of the chair, because of the course voxel size

2.5 Effects of Hyperparam variation

For mesh, the chamfer loss coefficient and the ico-sphere number has been varied keeping other parameters constant. Chamfer loss coefficient of 5.0 and ico-sphere value of 4 gave the best results. The table matches the logic that the higher chamfer loss and higher ico-sphere (more number of faces and vertices) leads to better reconstruction.

Chamfer loss weight	ico-sphere	F1 score
0.1	4	78.63
1	4	82.64
5	4	86.41
5	3	83.1
5	2	81.48

For point cloud, the number of point has been varied between 1000, 5000, 10000. All other parameters are constant. Num_point =10000, shows the best result because more number of points means higher granuality and better reconstruction

n_point	F1 score
1000	81.92
5000	88.27
10000	91.21

2.6

In this question, I wanted to analyze the rotation invariance of the model, especially because all the training data has been on the most usual chair/couch poses. In this experiment, I have rotated a couch, by 0,90,180 and 270 deg and predicted the point cloud representation. The point cloud representation is alll same for the 4 orientations, suggesting the model has overfitted to a single orientation. To overcome this, I would like to add rotations to the image and train the model again.