Tanzania: semantic_segmentation

Demo

Click on the below button to generate a random picture from Open AI Tanzania, and to display the corresponding label predictions.

Training procedure

We preprocessed the dataset images by subdividing raw images in 512*512-pixel tiles. The tiling process makes us getting around 90k training images and 3k validation images. The model used here is a U-net trained during 10 epochs, with validation phases.

Dataset description

The Open AI Tanzania dataset describes a building footprint recognition use case, where 3 types of buildings are discriminated: complete buildings, incomplete buildings and foundations. Doing semantic segmentation here means to define if a given pixel belongs to a building, and if required, the involved building type. The dataset contains 13 high-resolution images (6 to 8cm/pixel), amongst which 3 images are assigned to validation purpose. As a remark, there are 9 additionnal images without labelling information that may be used for testing models.