This python script helps you convert your mapillary vistas dataset to coco format.
Here is a given instance
image. Label info is embeded into each pixel value.
pixel / 256 # the value represents this pixel belongs to which label.
pixel % 256 # the value represents this pixel is the i-th instance of its label.
You can download the training and validation set in CoCo format (in JSON annotations) from the following links.
[1] Training Set
[2] Validation Set