Self-Supervised Learning for Stereo Reconstruction on Aerial Images
This work addresses the challenge of applying deep learning models for stereo estimation outside laboratory settings in aerial imaging, though it is incremental as it builds on existing hybrid CNN-CRF models with a self-supervised adaptation method.
The authors tackled the problem of requiring large amounts of labeled training data for stereo reconstruction in aerial images by proposing a self-supervised training procedure that adapts a network to dataset-specific characteristics without external ground truth, resulting in significant improvements in completeness and accuracy compared to a pre-trained model and favorable performance against other algorithms.
Recent developments established deep learning as an inevitable tool to boost the performance of dense matching and stereo estimation. On the downside, learning these networks requires a substantial amount of training data to be successful. Consequently, the application of these models outside of the laboratory is far from straight forward. In this work we propose a self-supervised training procedure that allows us to adapt our network to the specific (imaging) characteristics of the dataset at hand, without the requirement of external ground truth data. We instead generate interim training data by running our intermediate network on the whole dataset, followed by conservative outlier filtering. Bootstrapped from a pre-trained version of our hybrid CNN-CRF model, we alternate the generation of training data and network training. With this simple concept we are able to lift the completeness and accuracy of the pre-trained version significantly. We also show that our final model compares favorably to other popular stereo estimation algorithms on an aerial dataset.