ace19-dev/mvcnn-tf

Basic reference for Multi View Classification - mvcnn

MVCNN

Data

• Download modelnet10-Class Orientation-aligned Subset
  • You can create 2D dataset from 3D objects (.obj, .stl, and .off), using BlenderPhong
• Directly you get dataset https://drive.google.com/file/d/0B4v2jR3WsindMUE3N2xiLVpyLW8/view from https://github.com/RBirkeland/MVCNN-PyTorch/blob/master/README.md

Quick Start

• prepare multi-view image
• train.py

Evaluation

• eval.py

Retrieval

• deep metric learning
• extract features by train.py
• prepare query/gallery sources
• retrieve.py

TODO

• improve retrieval module.
• balanced sampling batch.

Notice

• The network architecture used in our experiments is relatively shallow to allow for fast training and inference.

References from

• https://github.com/WeiTang114/MVCNN-TensorFlow
• https://github.com/WeiTang114/BlenderPhong
• https://github.com/RBirkeland/MVCNN-PyTorch/blob/master/README.md
• https://github.com/tensorflow/models/tree/master/official/resnet
• https://arxiv.org/abs/1812.00442
• https://github.com/ace19-dev/gvcnn