This repository contains our final project for Machine Learning for 3D Geometry course at TUM. In this project we tried to improve the original workMulti-view Convolutional Neural Networks for 3D Shape Recognition by:
- Replacing backbone model from VGG to recent CNNs
- Introducing attention module
- Implementing a novel regularization method called ViewDropout which cuts information flow from some of the views in the training stage
The obtained results are given in the below table. We observed that adding attention mechanism did not improve the results but ViewDropout increased each models' metrics.
| Model Name | Baseline | Attention | Dropout |
|---|---|---|---|
| MVCNN-(Alexnet) | 90.64 | 90.44 | 91.17 |
| MVCNN-Resnet | 93.88 | 89.71 | 94.17 |
A sample output of our model:
The code for the original paper is based on RBirkeland's code and our implementation for model with the attention is very similar to the work Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction.
Sylvian Hofs