The ability to capture high-quality images under low-light conditions has been a long-standing pursuit within the computer vision community. Years of research has produced state-of-the-art algorithms exploiting techniques ranging from histogram equalization to retinex theory and, more recently, convolutional neural networks for low-light enhancement. Despite these advances, only a few of the existing methods can enhance extremely dark night-time images captured in near-zero lux conditions. Moreover, a practical solution must also respect additional constraints such as limited GPU memory and processing power with the remote device used for deployment. Existing methods, however, only target restoration quality and compromise on speed and memory requirements, raising concerns about their real-world deployability.

In this seminar, we present our extreme low-light night-time enhancement models that aim to strike a balance between network latency, memory utilization, model parameters, and reconstruction quality. To achieve this, we do most of the processing in the higher scale-spaces, skipping the intermediate-scales wherever possible. Also unique to our models is the potential to process all the scale-spaces concurrently, offering an additional 30% speedup without compromising the restoration quality. Our models can restore ultra-high-definition 4K resolution night-time images in just 1 sec. on a CPU and at 32 fps on a GPU while maintaining a competitive restoration quality.