Beyond Siamese KPConv: Early Encoding and Less Supervision

Iris de Gélis, Sébastien Lefèvre, Thomas Corpetti

Keywords: deep learning, change detection, point clouds

Abstract: Deep learning has led to significant advances in various fields, including Earth Observation, whose main tasks such as semantic segmentation or change detection have led to numerous deep network architectures. However, most of the deep networks for change detection still rely on 2D images (a.k.a. digital surface models), while the intrinsic nature of the changes is 3D. To the best of our knowledge, Siamese KPConv is the only deep model able to cope with 3D point clouds. We propose here to improve this model with early encoding of change information and less supervision.

Submission Number: 13