Transfer Mapillary Vistas Dataset to Coco format

This python script can help you transfer your MVD to coco format.

Currently we are working at MVD & ECCV 2018 competition, one of our plans is to transfer the dataset to coco format. And use currently prevalent frame implementation(mask RCNN) to train our model.

Brief Introduction

SplitTools.py finds label x instances tuple in each given instance image.

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.

Examples

Here we have a instance image.

we can split it according to label x instance relations.

1_41_0_Manhole is the location about the 0-th Manhole(Manhole's label_id is 41).

And then we use pycococreator tools to transfer them into json format, which is fit for coco data format.

I put the tools in Image2Json directory, and also you can download the tools here: Pycococreator tools