High-quality lighting positively influences visual performance in humans. The experienced visual performance can be measured using desktop luminance and hence several lighting control systems have been developed for its quantification. However, the measurement devices that are used to monitor the desktop luminance in existing lighting control systems are obtrusive to the users. As an alternative, ceiling-based luminance projection sensors are being used recently as these are unobtrusive and can capture the direct task area of a user. The positioning of these devices on the ceiling requires to estimate the desktop luminance in the user's vertical visual field, solely using ceiling-based measurements, to better predict the experienced visual performance of the user. For this purpose, we present LUMNET, an approach for estimating desktop luminance with deep models through utilizing supervised and self-supervised learning. Our model learns visual representations from ceiling-based images, which are collected in indoor spaces within the physical vicinity of the user to predict average desktop luminance as experienced in a real-life setting. We also propose a self-supervised contrastive method for pre-training LUMNET with unlabeled data and we demonstrate that the learned features are transferable onto a small labeled dataset which minimizes the requirement of costly data annotations. Likewise, we perform experiments on domain-specific datasets and show that our approach significantly improves over the baseline results from prior methods in estimating luminance, particularly in the low-data regime. LUMNET is an important step towards learning-based technique for luminance estimation and can be used for adaptive lighting control directly on-device thanks to its minimal computational footprint with an added benefit of preserving user's privacy.