Skip to content

Latest commit

 

History

History
73 lines (55 loc) · 2.23 KB

README.md

File metadata and controls

73 lines (55 loc) · 2.23 KB
Raw

3D Similar Voxel Data Search

This repository is for searching similar voxel data based on features compressed by 3D Convolutional Autoencoder.

Dependencies

tensorflow, scikit-learn, numpy, matplotlib

My test environment: Python3.6, tensorflow-gpu==1.12.0, scikit-learn==0.20.3, numpy==1.16.2, matplotlib==3.0.3

How to use

1. Prepare directory

Make sure that directories below exist.

  • log
  • checkpoint
  • data

2. Prepare data

Put your MODELNET10 data into data directory. The data should be npz form of array. The original data is available at:

PRINCETON MODELNET (http://modelnet.cs.princeton.edu/)

3. Train 3D CAE

Once array data is prepared, train 3D convolutional autoencoder and the checkpoint will be saved in checkpoint directory.

python train.py

You can also change training parameters by specifying arguments.

e.g.

pyhton train.py --num_epoch 100 --batch_size 32

Please look into the script about other settable parameters or run "python train.py --help"

4. Search similar data

Search similar data with the script based on trained model.

python evaluate.py

In this script, query data is randomly selected by "num_search_sample"(the number of query data) and similar data with numbers specified by "num_top_similarity" will be searched.
e.g. num_search_sample = 2, num_top_similarity = 3:
Two query samples and top three similar data by one query will be obtained.

Please look into the script about other settable parameters or run "python evaluate.py --help"

In the default setting, these result will be displayed with matplotlib. Depending on "num_search_sample" and "num_top_similarity", you might need long time to display.

Query samples

chair

dresser

toilet

Optional

You can also check distribution of encoded features by t-SNE and result of decoded data. Please activate comment lines in evaluate.py. (in defualt these are commented out)

# visualize encoded data with t-SNE
visualize_tsne(encoded, y_test)

# visualize input and its decoded data
visualize_3d_iodata(x_test[sample_idx], decoded[sample_idx], y_test[sample_idx])