This repository contains part segmentation experiments conducted on the ShapeNetPart dataset.
The ShapeNetPart dataset is annotated for 3D object part segmentation. It consists of 16,880 models from 16 shape categories, with 14,006 3D models for training and 2,874 for testing. The number of parts for each category is between 2 and 6, with 50 different parts in total.
pip install -r requirements.txtThe experiments are conducted on the following models:
Training models by running the corresponding scripts in the code folder. For example, to train the DGCNN model, run
the following command:
python code/train_dgcnn.pyThe processed dataset will be downloaded automatically when running the training scripts.
The table below presents the instance mIoU and class mIoU of the models on the ShapeNetPart dataset with 2048 points.
| Model | input | ins. mIoU | cls. mIoU | device |
|---|---|---|---|---|
| PointNet2SSG | xyz | 84.8% | 82.0% | 1x 3090 |
| PointNet2MSG | xyz | 85.2% | 82.5% | 1x 3090 |
| DGCNN | xyz | 85.4% | 83.1% | 1x 3090 |
You can reproduce the results by running the corresponding scripts in the code folder with default configurations.
For example, to train the PointNet2 model, run the following command
python code/train_pointnet2.py