Semantic Segmentation from Sparse Labeling using Multi-Level Superpixels

Implementation of our multi-level superpixel augmentation from sparse labeling presented on the IROS 2018.

Regarding our next work, CoralSeg: Learning Coral Segmentation from Sparse Annotations, we improve the algorithm regarding robustness and efficiency on memory (now it supports high resolution images) and speed (now it goes between x2 and x3 faster).

Citing Multi-Level Superpixels

If you find Multi-Level Superpixels useful in your research, please consider citing:

@inproceedings{alonso2019CoralSeg,
  title={CoralSeg: Learning Coral Segmentation from Sparse Annotations},
  author={Alonso, I{\~n}igo and Yuval, Matan and Eyal, Gal and Treibitz, Tali and Murillo, Ana C},
  booktitle={Journal of Field Robotics},
  year={2019}
}

@inproceedings{alonso2018MLSuperpixel,
  title={Semantic Segmentation from Sparse Labeling using Multi-Level Superpixels},
  author={Alonso, I{\~n}igo and Murillo, Ana C},
  booktitle={IEEE International Conference on Intelligent Robots and Systems (IROS)},
  year={2018}
}

Requirements

• Python 2.7
• OpenCV
• Numpy

Running it all

Generate a sparse ground-truth from a dense-labeled ground-truth

If your segmentation ground truth labels are dense, you can still simulate a sparse one, generating the sparse labeled images (images with only a few labeled pixels) with the following step.

The data should have this folder structure:

- dataset
	- images 
		- train
		- test
	- labels
		- train
		- test

Like the camvid dataset

To run this:

python generate_sparse/generate_sparse.py --dataset ./Datasets/camvid --n_labels 500  --gridlike 1 --image_format png --default_value 255

Every sparse labeled image will have [n_labels] number of labeled pixels. You can specify if you want the sparse labels to have a grid structure (value 1), or random (value 0) The rest of the pixels will have the [default_value] value.

The output folder will have the same name as the [dataset]: [dataset]/sparse_GT/train and [dataset]/sparse_GT/test

Generate augmented ground-truth

To generate the augmented ground-truth, you have to specify the path where the sparse labels and superpixels have been created.

python generate_augmented_GT/generate_augmented_GT.py --dataset ./Datasets/camvid --number_levels 15 --start_n_superpixels 3000 --last_n_superpixels 30

The output folder will have the same name as the [dataset]: [dataset]/augmented_GT/train and [dataset]/augmented_GT/test

Evaluation

For the evaluation of the quality of the augmented ground-truth, we compare it to the original dense labels. For this, you can run the following (specifying the dataset and the generated folder, as well as the number of classes to evaluate):

python evaluation/evaluate_augmentation.py --labels ./Datasets/camvid/ --generated ./camvid/ --classes 11

Training a semantic segmentation model

This other repository shows a simple example for training a semantic segmentation model using tensorflow with eager mode.