Capsule Nets for Content Based 3D Model Retrieval

First (as far as I know) investigation of the use of Capsule Networks for content-based 3d model retrieval. (information retrieval for 3D models)

This project was accomplished in just under 4 weeks. I probably used 120 hours of total compute time split between single GPU and 8 GPU aws instances (ModelNet40 takes FOREVER to train!). The final reported accuracies are Stratified 5-Fold cross validated.

Paper

Content-Based 3D Model Retrieval

Results

	ModelNet10		ModelNet40
	Score	2xStderr	score	2xStderr
Accuracy	93.08%	1.18%	82.73%	2.12%
Mean Average Precision	88.44%	4.15%	70.10%	2.22%

TODO:

• clean up code
• make replicable

*Cleaning up as we speak

this code has gotten pretty messy in the last week and I'm in the middle of cleaning it up. Please bare with me.

Setup

Download the data

cd 3d_model_retriever/
wget http://vision.princeton.edu/projects/2014/3DShapeNets/ModelNet10.zip
unzip ModelNet10.zip
wget http://modelnet.cs.princeton.edu/ModelNet40.zip
unzip ModelNet40.zip

Install binvox binary

In the root of the project

• MACOSX install:

    wget http://www.patrickmin.com/binvox/mac/binvox\?rnd\=1520896803665325 -O binvox
    chmod 755 binvox

• Linux install:

    wget http://www.patrickmin.com/binvox/linux64/binvox?rnd=1520896952313989 -O binvox
    chmod 755 binvox

Make env

cd 3d_model_retriever
python3 -m venv .env
source .env/bin/activate
pip install -r mac_requirements.txt

Convert all *.off files to *.binvox

python binvox_convert.py ModelNet10/ --remove-all-dupes
python binvox_convert.py ModelNet40/ --remove-all-dupes

Additional Notes

*.off files

Processing *.off files is done by a 3rd party tool made by Patrick Min.

• https://www.patrickmin.com/binvox/

What the hell is an *.off file?

• http://segeval.cs.princeton.edu/public/off_format.html

Voxel Resolution

• http://vision.princeton.edu/projects/2014/3DShapeNets/paper.pdf

"To study 3D shape representation, we propose to represent a geometric 3D shape as a probability distribution of binary variables on a 3D voxel grid. Each 3D mesh is represented as a binary tensor: 1 indicates the voxel is inside the mesh surface, and 0 indicates the voxel is outside the mesh (i.e., it is empty space). The grid size in our experiments is 30 × 30 × 30."

./binvox -cb -e -c -d 30 sample.off

• -e is important, with a lot of troubleshooting it was shown that not using this led to inconsistent voxelization :headache:

** Viewing .binvox data for troubleshooting **

./viewvox <filename>.binvox

Load The Data in Python

from data import load_data
(x_train, y_train), (x_test, y_test), target_names = load_data('./ModelNet10')