AIRI_Pottery

3D classification and segmentation (location prediction) of points cloud data

• Each pottery(label) has various size of shards.
• Each Shards has various size of points.

---

DATA PREPARING

Data원본: AIRI-NAS(http://airi-nas.local:5000/)/Shared/Pottery/pottery_full_data 에 있습니다.

실제 학습에 필요한 point cloud 데이터는 pottery_full_data의 하위폴더들에 있는 .npy 파일들이며 각 label별로 폴더에 따로 저장해서 사용하시면 됩니다.(make_filelist.ipynb 참고)

1. run make_filelist.ipynb
   • 도자기 별 shards list 생성(filelist_라벨_파편수_randomseed_F.txt)
2. run edit_h5_seglabel.ipynb
   • 도자기 별 shards' CENTRALIZED point cloud data, pottery labels(ID), segmentation_label 생성(라벨_파편수_randomseed.h5),
   • segmentation_label: 각 shard의 (y_max, y_mean, y_min) 을 label로 부여
   • train data, test data list 생성(train_files.txt, test_files.txt)
3. 위에서 생성한 h5 형식의 data와 train_files.txt, test_files.txt 를 한 폴더로 이동(provider.py 와 train_pottery_combined.py 에서 경로 참조)

---

Model

• train: ./train_pottery_combined.py
• model: ./models/dgcnn+skipdense.py

python train_pottery_combined.py

---

RESULT FIGURE (ex)

1. classification

2. location prediction

---

EXPERIMENT

• 실험 환경: ubuntu 16.04, 64Gmemory, 16core, GPU Tesla V100-SXM2(16G) * 2 (used 1 GPU)
• Tensorflow

---

VISUALIZATION

./pottery_demo.ipynb

• required: tetgen library

$ conda insatll -c conda-forge tetgen

---

DEMODAY VIDEO

./demoday_pottery.mkv
./demoday_pottery.mp4

Acknowledgement

This code is heavily borrowed from dgcnn, and pointnet