Learning Aerial Image Segmentation from Online Maps

What they do

• Pixel-accurate, but small-scale ground truth available.
• Less accurate reference data that is readily available in arbitrary quantities, at no cost.

Tools

• OpenStreetMap (OSM)

Description

They use OSM like weakly labeled training data for three classes, buildings, roads and background with RGB orthophotos from Google Maps.

The author’s hypotheses

• The volume of training data can possibly compensate for lower accuracy of the labeling.
• The large variety present in very large training set could potentially improve the classifier’s ability to generalize to new unseen location.
• The addition of the large volume of weak data to low volume high-quality data could potentially improve the classification.
• If low-accuracy, large-scale training data helps, it may allow substituting a large portion of the manually annotated high-quality data.

Model

FCN (Fully convolutional network) with modifications and as input, an 500x500pixel patch (mini batch 1 image).

They make some pre-training before the real training to get better results. One by the Pascal VOC 2010¹ and the other with the OSM data.

Experiments

• Complete substitution (No manual labeling).
• Augmentation (pre-training help or not).
• Partial substitution (pre-training with OSM labeling in the training).

Datasets

ground sampling distance (GSD).

Potsdam - 1 image (6000x6000 pixel) = 144 images (500x500 pixel), only 21 have been manually labeled.

Results

For Chicago with only the OSM training.

The methode 5² is:

The finals results are:

Conclusions

• Large scale, nevertheless significantly improves segmentation performance, and improves generalization ability of the models.
• Training only on open data, without manual labelling, achieves reasonable results.
• Large-scale pre-training with OSM labels significantly benefits semantic segmentation.

---