markus-hinsche/weakly-supervised-semseg

Weakly Supervised Semantic Segmentation

Setup

Local

Create new python virtual environment. Then:

pip install -r requirements.txt

Cloud (GCP)

./gcp_instance.sh

Data

Find the data here: ​ISPRS Vaihingen dataset​

There are 33 high-resolution (approx 1500x2000px) aerial images of the town of Vaihingen, Germany. Labels are supplied with 1 class per pixel, as follows:

1. Impervious surfaces - WHITE
2. Building - BLUE
3. Low vegetation - TURQUOISE
4. Tree - GREEN
5. Car - YELLOW
6. Clutter/background - RED

Data preprocessing

src/split_to_tiles.py: Cuts 33 high-resultion images into 4497 200x200 tiles.

src/prep_tiles.py: Creates weakly-supervised tile-level annotations.

src/split_sets.py: Creates the N1, N2, N_validation split (randomly), see src/constants.py for the resulting split.

src/mask_conversion.py: Takes 3 channel RGB masks and translates them into 1 channel images where a pixel value is encoded by 0, 1, 2, 3, ..., n_colors

Project

Weakly supervised learning with some fully supervised (pixel-level) annotations.

src/fully-supervised-semseg.ipynb: Fully-supervised (FS) training on N1

src/weakly-supervised-semseg.ipynb: Weakly-supervised (WS) training on N2

src/mixed-supervision-semseg.ipynb: Combine FS and WS to try to improve the performance of the semantic segmentation task.

Report and Results

See <report.md>

Run unit tests

pytest