**Assignment 2**
Student name: Gaurav Parmar
(#) 1. Exploring Loss Functions
(##) 1.1 Fitting a voxel grid
The ground truth target (left) and optimized (right) voxels
(##) 1.2 Fitting a point cloud
The ground truth target (left) and optimized (right) point clouds
(##) 1.3 Fitting a mesh
The ground truth target (left) and optimized (right) meshes
(#) 2. Reconstructing 3D from single view
(##) 2.1 Image to voxel grid
The input image (left), ground truth voxel grid (mid), predicted voxel grids (right) shown below.
(##) 2.2 Image to point cloud
The input image (left), ground truth mesh (mid), predicted point clouds (right) shown below.
(##) 2.3 Image to meshes
The input image (left), ground truth mesh (mid), predicted mesh (right) shown below.
(##) 2.4 Quantitative Comparison
Avg F1@0.05 for voxels: 67.63
Avg F1@0.05 for point clouds: 91.57
Avg F1@0.05 for meshes: 83.73
The performance is the worst for voxel grids and the best for the point clouds.
The relatively bad performance of the voxel representations is because of the
low spatial resolution used with voxels.
(##) 2.5 Analyse effects of hyperparms variations
First, I vary the number of points used when using the point cloud representations.
Note that I use fewer traiing steps than the previous experiments for these ablations.
Avg F1@0.05 for num=100: 42.9
Avg F1@0.05 for num=1k: 78.1
Avg F1@0.05 for num=10k: 88.7
As the number of points is increased, the score increases monotonically as point complex
and high frequency components of the scene can be represented.
(##) 2.6 Interpret your model
(###) I visualize what are type of images our model performs the best and worst on.
Images/grids with the highest F1 score
Images/grids with the lowest F1 score
Observations
From these visualizations, I can unsurprisingly observe that the stereotypical chairs that have
four legs and a backrest are easy to convert to 3D. However images shown in the second table
that are a bowling ball, a stool with a circular legs with holes, and a square table with no
backrest are understandly difficult to reconstruct and have a low F1 score.
(#) late days
I regretably used 5 late days on this assigment.
