Corinne Alini HW 2
I have used 0 late days on this assignment
1.1. Fitting a voxel grid
Target, Result
1.2. Fitting a point cloud
Target, Result
1.3. Fitting a mesh
2.1: Image to voxel grid
Example 1
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 2
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 3
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Question 2.2: Image to Point Cloud
Example 1
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 2
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 3
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
2.3: Image to Mesh
Example 1
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 2
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Example 3
Input RGB ----------- Ground Truth Voxel-----------Predicted Voxel
Question 2.4: Quantitative Comparisons
Vox: 63.001
Points:93.772
Mesh: 86.324
Point cloud has the best f1 score while vox has the worst f1 score. This intuitively makes sense because
the point cloud only has to learn x,y,z pairs and thus does not have to learn surface or connectivity. However,
mesh and point cloud have very similar f1 scores because they both are learning an offset. This is why they have
similar loss functions.
The reason that voxel has the worst f1 score is it is learning occupancy probabilities then we are having to convert
it to a mesh to compare it to a mesh. We are not comparing it directly to a voxel. This adds a layer of complexity that
would make the comparisons between mesh vs voxel worse than, say for example a mesh vs mesh comparison.
Question 2.5:Analyze effects to hyperparameter variations
I decided to focus on the laplacian smoothing parameter. I am changing the ratio of chamfer loss vs smoothing.
When I increased smoothing, we are weighting the smoothing higher; it removed some details that are important features of the chair.
When I decreased smoothing, we weighted smoothing low. The features of the image were more pronounced but were very jagged.
There were features of the chair that were not part of the ground truth mesh. There was a tradeoff between making the mesh
look unrealistically smooth by removing important features of the chair and decreasing features that were important to chair identification.
Question 2.6: Interpret your model
I chose to make a visualization of how the model trains. At every 50 iterations, I ran the same
example image through the model and outputted an image of what the model predicts at the given time in training.
I used mesh as an example visualization. You can see that the model deforms the image by calculating the offset of the
original mesh. This representation is useful as it highlights how the deformation works while training a mesh.
Example of voxel training over time
Example of point cloud training over time
