Bayesian nights: Optimizing night photography rendering with Bayesian derivative-free methods

We introduce a novel approach for optimizing Image Signal Processing (ISP) rendering pipelines for night photography through a Bayesian derivative-free procedure. Traditional neural-network-based ISPs depend on differentiable operations to enable backpropagation-based optimization, a requirement that can impose significant constraints. Our method circumvents this by employing Bayesian optimization to fine-tune the pipeline's parameters, independently of their differentiability. Additionally, we address the need for paired data to enable supervised optimization: while such paired data is available on public datasets, it is expensive to collect for new imaging devices. To this extent, we design a raw-to-raw mapping procedure, that aligns images from an available paired dataset to the target unpaired dataset. This allows us to supervise the optimization of our solution directly within the target space, without the need for device-specific paired data. We validate our approach with extensive experimentation on paired and unpaired datasets, demonstrating its efficacy using both subjective and objective evaluation metrics. Our code is made available for public download at https://github.com/TheZino/Bayesian-pipeline-optimization.