SeedFormer: Patch Seeds based Point Cloud Completion with Upsample Transformer

This repository contains PyTorch implementation for SeedFormer: Patch Seeds based Point Cloud Completion with Upsample Transformer (ECCV 2022).

SeedFormer presents a novel method for Point Cloud Completion. In this work, we introduce a new representation, namely Patch Seeds, which not only captures general structures from partial inputs but also preserves distinctive information of local patterns. Moreover, we devise a novel Upsample Transformer by extending the transformer structure into basic operations of generating points, which explicitly incorporates spatial and semantic relationships in the local neighborhood.

🔥Updates

• 2022-07-05: Initial Update.

Installation

The code has been tested on one configuration:

• python == 3.6.8
• PyTorch == 1.8.1
• CUDA == 10.2
• numpy
• open3d

pip install -r requirements.txt

Compile the C++ extension modules:

sh install.sh

Datasets

Download dataset files from here.

The details of used datasets can be found in DATASET.md (thank the authors of PoinTr).

Usage

Training on PCN dataset

First, you should specify your dataset directories in train_pcn.py:

__C.DATASETS.SHAPENET.PARTIAL_POINTS_PATH        = '<*PATH-TO-YOUR-DATASET*>/PCN/%s/partial/%s/%s/%02d.pcd'
__C.DATASETS.SHAPENET.COMPLETE_POINTS_PATH       = '<*PATH-TO-YOUR-DATASET*>/PCN/%s/complete/%s/%s.pcd'

SeedFormer takes two V100 gpus with a batch size of 48. To train SeedFormer on PCN dataset, simply run:

python3 train_pcn.py

Testing on PCN dataset

To test a pretrained model, run:

python3 train_pcn.py --test

Or you can give the model directory name to test one particular model:

python3 train_pcn.py --test --pretrained train_pcn_Log_2022_XX_XX_XX_XX_XX

Save generated complete point clouds as well as gt and partial clouds in testing:

python3 train_pcn.py --test --output 1

Acknowledgement

Some parts of the code are borrowed from GRNet and SnowflakeNet. We thank the authors for their excellent work.