How to train data

OpenSSE offers you a convenient command sse after installed.

Database

• Database in my demo comes from SHREC 2012, download 3D target dataset(~112MB). We use models under Watertight_dataset/Extended folder.

  $ mv ~/Download/Watertight_dataset/Extended ~/Database/SHREC12/

  Notice: Currently download link is unavailable, so you can download dataset from opensse-demo docker images.

• Or download SHREC 2012 3D target dataset from opensse-demo docker images as below steps:

  1. Download docker image

  $ docker pull zddhub/opensse-demo

  2. Create docker container named opensse-demo

  $ docker run -it --name opensse-demo zddhub/opensse-demo /bin/bash

  3. Create a new terminal and copy dataset from docker containers

  docker cp opensse-demo:/Users/zdd/Database/SHREC12 ~/Database/

• Or you also can use your own models.

Train data

Get line drawing views

Using my another project, you will get 102 views for each model. Please visit this link for details.

Generate vocabulary

• Step 1: Generate line drawing images filelist

$ sse filelist -d ~/Database/SHREC12/ -p "*.jpg" -o filelist

• Step 2: Extract descriptors

$ sse extract -f filelist -o features

• Step 3: Generate vocabulary (optional)

$ sse vocabulary -f features -n 1000 -o vocabulary

It will take you some time, but only need run this once. or directly use my vocabulary under ../opensse/data/.

Create inverted index file

• Step 4: Quantize feature

$ sse quantize -v vocabulary -f features -o samples

When dealing with large scale database, we can directly generate samples using extract_and_quantize command to reduce disk usage without feature files.

$ sse extract_and_quantize -f filelist -v vocabulary -o samples

• Step 5: Create inverted index file

$ sse index -s samples -o index_file

Now all data are ready. Go to Play with Command Line and Play with SketchSearchDemo.