An unofficial PyTorch implementation of ICCV 2019 paper "Orientation-Aware Semantic Segmentation on Icosahedron Spheres". Only HexRUNet-C for Omni-MNIST is implemented right now.
Python 3.6 or later is required.
Python libraries:
- PyTorch >= 1.3.1
- torchvision
- tensorboard
- tqdm
- igl
Run the following command to train with random-rotated training data and evaluate with random-rotated test data.
python train.py --train_rot --test_rotYou can change parameters by arguments (-h option for details).
Here is the results of this repository. Accuracy of the last epoch (30th epoch) is reported.
Omni-MNIST HexRUNet-C accuracy (%)
| N/N | N/R | R/R | |
|---|---|---|---|
| This repository | 99.15 | 69.62 | 98.36 |
| Paper | 99.45 | 29.84 | 97.05 |
N/N: Non-rotated training and test dataN/R: Non-rotated training data and random-rotated test dataR/R: Random-rotated training and test data
As can be observed here, N/R of this repogitory is much higher than the one reported in original paper. I guess it's because the implementation of projecting images on a sphere and rotation are different (My implementation of the projection is based on ChiWeiHsiao/SphereNet-pytorch).