Using the inverse square law we are able to estimate the depth and albedo from a single image using deep learning without supervision. (2/7)

We exploit the fact that in endoscopies there is a single light source. Thus, one can clearly see that the darker regions correspond to the distant points. (3/7)

In this work, we train a neural network to predict the depth and albedo of the image. Then, we compute the surface normals from the depth map which allows us to use a differentiable rendering equation as a self-supervised loss. (4/7)

We are able to achieve performance close to fully supervised learning, without using any label. Because we are using a self-supervised loss, we are also able to perform test-time refinement (TTR). (5/7)