akorol98 / vaihingen-semantic-segmentation

Vaihingen-Semantic-Segmentation

Outline:

• Set Up Conda Environment

• Data Gathering and Preprocesing

• Model training

Set Up Conda Environment

conda create -f environment.yml

Data Gathering and Preprocesing

Step 1: ISPRS Vaihingen dataset

The ISPRS Vaihingen dataset can be downloaded with the login details in an automated email, after completing this form form via:

ground truth:

wget ftp://$username:[email protected]/ISPRS_BENCHMARK_DATASETS/Vaih
ingen/ISPRS_semantic_labeling_Vaihingen_ground_truth_COMPLETE.zip

image data:

wget ftp://$username:[email protected]/ISPRS_BENCHMARK_DATASETS/Vaih
ingen/ISPRS_semantic_labeling_Vaihingen.zip

Ones the data has been downloaded move the archives to the data folder in project directory and unzip them:

unar -d ISPRS_semantic_labeling_Vaihingen.zip

unar -d ISPRS_semantic_labeling_Vaihingen_ground_truth_COMPLETE.zip

Step 2: Preprocesing

python preprocessing.py

Model training

Baseline Model

To train a baseline model, run:

python train.py baseline

Weakly Supervised Model

The weakly supervised training is performed in three steps:

Step 1: Train baseline

Train the baseline model (if still not trained). For that see the Baseline Model section.

Step 2: Generate masks for weakly training

Generate the masks using the trained baseline model. For this run from the main challenge directory

python gen_masks.py

This will generate the masks which are corrected by image level labels: if on the image level some tag is not present, all the corresponding pixel level tags of this type are deleted.

Step 3: Run the actual training of the weakly supervised model

python train.py weakly_supervised

The idea behind this model is that one uses the masks generated by the baseline model to use this information as a support to training on N1 + N2(weakly) combined sample.