Part Segmentation on ShapeNetPart

Note in this experiment, we work on the ShapeNetPart Segmentation. The number of parts for each category is between 2 and 6, with 50 different parts in total. The data we use is: shapenetcore_partanno_segmentation_benchmark_v0_normal.zip[1]. We uniformly sample 2048 points in training and testing.

Dataset

Download the dataset from the official website, put the data under data/ShapeNetPart/, and then run the training code once. The data will be autmmatically preprocessed (uniformly sample 2048 points).

You can also use our preprocessed data provided below:

cd data && mkdir ShapeNetPart && cd ShapeNetPart
gdown https://drive.google.com/uc?id=1W3SEE-dY1sxvlECcOwWSDYemwHEUbJIS
tar -xvf shapenetcore_partanno_segmentation_benchmark_v0_normal.tar

Organize the dataset as follows:

data
 |--- ShapeNetPart
        |--- shapenetcore_partanno_segmentation_benchmark_v0_normal
                |--- train_test_split
                      |--- shuffled_train_file_list.json
                      |--- ...
                |--- 02691156
                      |--- 1a04e3eab45ca15dd86060f189eb133.txt
                      |--- ...               
                |--- 02773838
                |--- synsetoffset2category.txt
                |--- processed
                        |--- trainval_2048_fps.pkl
                        |--- test_2048_fps.pkl

Train

For example, train PointNeXt-S using 4 GPUs by default

CUDA_VISIBLE_DEVICES=0,1,2,3 python examples/shapenetpart/main.py --cfg cfgs/shapenetpart/pointnext-s.yaml

- change cfg to cfgs/shapenetpart/pointnext-s_c160.yaml to train the best model we report in our paper.

Test

CUDA_VISIBLE_DEVICES=0 python examples/shapenetpart/main.py --cfg cfgs/shapenetpart/pointnext-s.yaml mode=test --pretrained_path /path/to/your/pretrained_model

Profile parameters, FLOPs, and Throughput

CUDA_VISIBLE_DEVICES=0 python examples/profile.py --cfg cfgs/shapenetpart/pointnext-s.yaml batch_size=64 num_points=2048 timing=True flops=True

Reference

@article{yi2016shapnetpart,
Author = {Li Yi and Vladimir G. Kim and Duygu Ceylan and I-Chao Shen and Mengyan Yan and Hao Su and Cewu Lu and Qixing Huang and Alla Sheffer and Leonidas Guibas},
Journal = {SIGGRAPH Asia},
Title = {A Scalable Active Framework for Region Annotation in 3D Shape Collections},
Year = {2016}}`