Physics-based vision meets deep learning

Physics-based vision explores computer vision and graphics problems by applying methods based upon physical models. On the other hand, deep learning is a learning-based technique, where a substantial number of observations are used to train an expressive yet unexplainable neural network model. In this thesis, we propose the concept of a model-based decoder, which is an unlearnable and differentiable neural layer being designed according to a physics-based model. Constructing neural networks with such model-based decoders afford the model strong learning capability as well as the potential to respect the underlying physics.

We start the study by developing a toolbox of differentiable photometric layers ported from classical photometric techniques. This enables us to perform the image formation process given geometry, illumination and reflectance function. Applying these differentiable photometric layers into a bidirectional reflectance distribution function (BRDF) estimation network training, we show the network could be trained in a self-supervised manner without the knowledge of ground truth BRDFs.

Next, in a more general setting, we attempt to solve inverse rendering problems in a self-supervised fashion by making use of model-based decoders. Here, an inverse rendering network decomposes a single image into normal and diffuse albedo map and illumination. In order to achieve self-supervised training, we draw inspiration from multiview stereo (MVS) and employ a Lambertian model and a cross-projection MVS model to generate model-based supervisory signals.

Finally, we seek potential hybrids of a neural decoder and a model-based decoder on a pair of practical problems: image relighting, and fine-scale depth prediction and novel view synthesis. In contrast to using model-based decoders to only supervise the training, the model-based decoder in our hybrid model serves to disentangle the intricate problem into a set of physically connected solvable ones. In practice, we develop a hybrid model that can estimate a fine-scale depth map and generate novel view synthesis from a single image by using a physical subnet to combine results from an inverse rendering network with a monodepth prediction network. As for neural image relighting, we propose another hybrid model using a Lambertian renderer to generate initial estimates of relighting results followed by a neural renderer performing corrections over deficits in initial renderings.
We demonstrate the model-based decoder can significantly improve the quality of results and relax the demands for labelled data.