This thesis work explores self-supervised learning methods based on motion in videos to reduce the reliance on costly annotated datasets for the tasks of optical flow and monocular depth estimation. In the absence of ground truth, both tasks are mainly learned with an image reconstruction loss, which relies on the brightness constancy hypothesis.
In practice, this assumption may not be verified due to brightness changes often caused by moving shadows or non-Lambertian surfaces, which prevents some reconstructions. On the one hand, solutions can be implemented to limit the impact of these brightness changes. Thus, our first contribution improves the performance of self-supervised optical flow estimation methods thanks to a neural network designed to compensate for any brightness change at the training only, so that the running time at inference is not affected. On the other hand, since the reconstruction loss limits make some cases poorly supervised and therefore difficult to estimate for a depth estimation neural network, they are a source of aleatoric uncertainty that can be estimated. In our second contribution, we show that using our new probabilistic formulation of the problem of self-supervised learning of monocular depth provides both better depth and uncertainty predictions.