EPiC: Ensemble of Partial Point Clouds for Robust Classification (ICCV 2023)

Official pytorch implementation of the paper "EPiC: Ensemble of Partial Point Clouds for Robust Classification".

Meir Yossef Levi, Guy Gilboa

We propose a novel approach to robustify performance of point-cloud classification networks against corruptions and outliers. Our method is based on ensemble of 3 types of sub-samples: Patches, Curves and Random. It can be applied to any given point-cloud network and achieve state-of-the-art results on ModelNet-C by using RPC+WolfMix (mCE=0.501).

---

Tested on Ubuntu 20.04 LTS using python 3.8.10

EPiC notebook:

Get Started

Step 0. Clone the Repo

git clone https://github.com/yossilevii100/EPiC.git
cd EPiC

Step 1. Set Up the Environment

Set up the environment by:

pip install pointnet2_ops_lib/.

Step 2. Prepare Data

Download ModelNet-40 by:

cd data
wget https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip
unzip modelnet40_ply_hdf5_2048.zip
cd ..

Alternatively, you may download ModelNet40 manually and extract it under data.

Download ModelNet-C by:

cd data
gdown https://drive.google.com/uc?id=1KE6MmXMtfu_mgxg4qLPdEwVD5As8B0rm
unzip modelnet_c.zip
cd ..

Alternatively, you may download ModelNet40-C manually and extract it under data.

Make sure your data folders are as follows:

data
--- modelnet_c
------ list of h5 files
--- modelnet40_ply_hdf5_2048
------ list of h5 files

Step 3. Download Pretrained Models

Download pretrained models by

gdown https://drive.google.com/uc?id=15Q-YewNGvte8PmteVjTmqE0vzDL3ViJd
unzip pretrained.zip -d pretrained

Alternatively, you may download pretrained models manually and extract it under root directory.

Evaluation Commands

Evaluation commands are provided in EVALUATE.md.

Customize EPiC training+evaluation for your custom model

In order to ease the way for future researchers we implement a placeholder for new custom model. All you have to do is implement your custom model in models/custom/custom_model

• Augmented by WolfMix:

Train:

python main.py --model custom_model --train_random --train_curves --train_patches --use_wolfmix --exp_name <your_exp_name> --use_wolfmix

After the training procedure finished, evaluate your model with EPiC by:

python main.py --model custom_model --eval --model_path_patches <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_patches_wm.t7 --model_path_curves <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_curves_wm.t7 --model_path_random <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_random_wm.t7

• Un-augmented:

Train:

python main.py --model custom_model --train_random --train_curves --train_patches --exp_name <your_exp_name>

After the training procedure finished, evaluate your model with EPiC by:

python main.py --model custom_model --eval --model_path_patches <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_patches.t7 --model_path_curves <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_curves.t7 --model_path_random <path/to/project>/checkpoints/<your_exp_name>/models/custom_model_random.t7

Results on ModelNet-C

Our method achieves SOTA results on ModelNet-C, with and without augmentation. Moreover, our approach is improved each of the examined networks in terms of robustness.

With WolfMix Augmentation

Method	Reference	mCE	Clean OA
DGCNN	Wang et al.	0.590	0.932
DGCNN-EPiC	Wang et al.	0.529	0.921
---------------	----------------------------------------------------------	:---:	:------:
GDANet	Xu et al.	0.571	0.934
GDANet-EPiC	Xu et al.	0.530	0.925
---------------	----------------------------------------------------------	:---:	:------:
PCT	Guo et al.	0.574	0.934
PCT-EPiC	Guo et al.	0.510	0.927
---------------	----------------------------------------------------------	:---:	:------:
RPC	Ren et al.	0.601	0.933
RPC-EPiC	Ren et al.	0.501	0.927

Without WolfMix Augmentation

Method	Reference	mCE	Clean OA
DGCNN	Wang et al.	1.000	0.926
DGCNN-EPiC	Wang et al.	0.669	0.930
---------------	----------------------------------------------------------	:---:	:------:
GDANet	Xu et al.	0.892	0.934
GDANet-EPiC	Xu et al.	0.704	0.936
---------------	----------------------------------------------------------	:---:	:------:
PCT	Guo et al.	0.925	0.930
PCT-EPiC	Guo et al.	0.646	0.930
---------------	----------------------------------------------------------	:---:	:------:
RPC	Ren et al.	0.863	0.930
RPC-EPiC	Ren et al.	0.750	0.936
---------------	----------------------------------------------------------	:---:	:------:
CurveNet	Xiang et al.	0.927	0.938
CurveNet-EPiC	Xiang et al.	0.742	0.921

Cite EPiC

@inproceedings{levi2023epic,
  title={Epic: Ensemble of partial point clouds for robust classification},
  author={Levi, Meir Yossef and Gilboa, Guy},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={14475--14484},
  year={2023}
}

Acknowledgement

This codebase heavily borrows codes from the following repositories:

• ModelNet-C
• DGCNN
• PCT_Pytorch
• GDANet
• CurveNet
• RSMix
• PointWOLF