Skip to content

hjsg1010/pottery

Repository files navigation

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.

We aim to classify shards into different pottery classes and predict a relative location of shards within the pottery using point cloud representation.

abstract_figure


Data Preparing

We used 3D scanned pottery data in .npy format.

The figure below is an example of the pottery prototype and shards that we actually used. original_pottery_shards

Also, we first generated deep learning models to generate synthetic data using different methods as shown in below figure data_generation

Data Preprocessing

  1. run make_filelist.ipynb
    • generated shards list per each pottery (filelist_label_numofshards_randomseed_F.txt)
  2. run edit_h5_seglabel.ipynb
    • shards' CENTRALIZED point cloud data per pottery, pottery labels(ID), segmentation_label 생성(label_numofshards_randomseed.h5),
    • segmentation_label: labelling (y_max, y_mean, y_min) of each shard
    • generate train data, test data list (train_files.txt, test_files.txt)
  3. move h5 format data, train_files.txt and test_files.txt into same folder (please refer the path in provider.py and train_pottery_combined.py)

Model

model_figure

- Simultaneous learning of pottery type classification and relative position prediction
  • train: ./train_pottery_combined.py
  • model: ./models/dgcnn+skipdense.py
python train_pottery_combined.py

Result Figure (ex)

1. Classification

classification

2. Location Prediction

segmentation


Experiment

  • Experiment settings: ubuntu 16.04, 64Gmemory, 16core, GPU Tesla V100-SXM2(16G) (used 1 GPU)
  • Tensorflow

Visualization

run pottery_demo.ipynb

  • required: tetgen library
$ conda insatll -c conda-forge tetgen

Acknowledgement

This code is heavily borrowed from dgcnn, and pointnet