Aerial: semantic_segmentation

Demo

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

Training procedure

We preprocessed the dataset images by subdividing raw images in 250*250 tiles, that are resized so as to get 224*224 images. The tiling process makes us getting 66000 training images, 6000 validation images, and 72000 testing images (each original images gives 400 preprocessed images). The model used here is a U-net trained during 2 epochs, with validation phases.

Dataset description

The Aerial image dataset contains only 2 labels, i.e. building or background and consequently the model aims at answering one single question for each image pixel: does this pixel belongs to a building? The dataset contains 360 images, one half for training one half for testing. Each of these images are 5000*5000 tif images. Amongst the 180 training images, we assigned 15 training images to validation.