sorenmulli / imgseg-fake-balloons

Synthetic Semantic Segmentation Training Data using WebGL

Course project for 02561 Computer Graphics at the Technical University of Denmark.

See report and the example website.

Reproduction

Setup

Run make to install the Python software. Open webgl-site/index.html in any modern browser to see the examples.

Get Fake Balloons data

You can build the synthetic dataset yourself; see make data. The version used in this project is downloadable from Google Drive here and here or by running wget commands from the script dev/download.sh.

Get COCO data

Download 2017 test and val images and annotations from COCO download page. Place in a directory, e.g. cocodata where structure should be

cocodata/
├─ instances_train2017.json
├─ instances_val2017.json
├─ val2017/
│  ├─ 000000000139.jpg
│  ├─ ...
├─ train2017/
│  ├─ 000000000009.jpg
│  ├─ ...

Run experiments

See python imgseg02561/train.py --help. My configuration is run when running make train but some paths should be configured to your data locations.