Demo

Click on the below button to generate a random picture from Mapillary dataset, and to display the corresponding label predictions.

Generate a new image

Training procedure

A U-net model has been trained during 20 epochs with a set of 18000 training images, with a validation procedure involving 2000 images at the end of each training epoch.

Dataset description

The original Mapillary dataset contains 65 labels. For a sake of clarity, a lighter classification with only 13 labels has been built up:

• animal: ground animal or birds
• construction (barrier): curbs, fences, guard rails and other barriers
• construction (flat): ground areas like roads, parkings, bike lanes...
• construction (structure): buildings, bridges, tunnels
• human
• rider: human on a bike, a motorbike or a horse
• marking
• nature: mountain, sand, sky, vegetation
• traffic-sign
• support: pole, traffic-sign frame, utility pole
• vehicule
• object: miscellaneous objects
• void: own vehicle or unlabeled pixels