Try to unstuck: back to COCO

The last post is from October 2020. The main line of conduct was to progress on chromosome instance segmentation, but a robust semantic, U-net based, would have satisfying too (possibly using Fastai).

Detectron2, PixelLib and many others provide instance segmentation algorithms (Mask RCNN for example). To train a model, the COCO format for the so called ground-truth labels seems to be mandatory. The issue is that the different datasets generated to simulate overlapping chromosomes, the labels are grey scaled images decomposable into binary masks for one-hot encoding:

 

In the last try, the idea was to start from the COCO specs and to write some code to convert the binary masks into COCO files but that was a fail as detectron2 didn't want my minimalist dataset.

Making a minimal valid coco data

From a grey-scaled image a coco file is generated using an interactive online tool as https://www.makesense.ai/:

 

The coco dataset corresponding to this single image is a json file :

 

With a xml viewer in Colab, we can see how the file is structured:

The file corresponds to only one image:"grey0000001.png"

The two chromosomes annotated appeared as id:0 and id:0 in the annotations field:

The contour of one of the two chromosomes is coded a 24 values. Possibly 12 pairs of coordinates:

The chromosome bounding box seems to be defined by two diagonal points, so we have a pair of coordinates:

Finally, there are only one category of instances: "chromosome"

Back to COCO and play with a minimalist valid dataset with pycocotools:

COCO dataset from scratch : try and fail ...

Detectron2 provides several algorithms for instance segmentation, so it was tempting to submit the overlapping datasets to one of those. However, to use one of these algorithms, the dataset format seem to follow the MS-COCO format.

One available dataset consists in 2164 pairs of grayscaled+groundtruth images.To give a try, a minimalist dataset with one image and two instances could be converted to COCO format:

 

The two instances (right) are obtained from the groundtruth image showing the overlapping chromosomes. The instances are numpy arrays which can be saved as png images. To generate a COCO dataset associated to the gray scaled image (left), the following steps were followed:

• generate a python dictionary according to the COCO format specification found in the detectron2 documentation and convert the binary masks to their bounding boxes and compressed rle using pycocotools.
  
• Save the dictionary as a json file
• Load the json file with pycocotools (or detectron2) in order to visualize if possible the instances overlaying the gray scaled image.

 The whole process is available in a jupyter notebook on Kaggle.

Unfortunately, the dataset is not a legit COCO dataset as the dataset registration fails. Hope to get some help on Pytorch forum or from stackoverflow.