The paper addresses the challenge of imaging in extremely low light conditions, where traditional camera processing pipelines fail due to low photon counts and low signal-to-noise ratio (SNR). The authors introduce a new dataset, the See-in-the-Dark (SID) dataset, which contains 5094 raw short-exposure images paired with corresponding long-exposure reference images. This dataset is designed to support the development of data-driven approaches for low-light image processing. To address the limitations of traditional methods, the authors propose a pipeline based on end-to-end training of a fully-convolutional network (FCN). The FCN processes raw sensor data directly, bypassing the traditional image processing pipeline, which often performs poorly in such conditions. The network is trained to perform tasks such as color transformations, demosaicing, noise reduction, and image enhancement. The SID dataset and the proposed pipeline are evaluated through various experiments, including qualitative comparisons with traditional pipelines, denoising algorithms, and burst processing. The results show that the proposed pipeline significantly outperforms existing methods in terms of noise suppression and color transformation. The authors also conduct perceptual experiments using Amazon Mechanical Turk to assess user preference, finding that their pipeline significantly outperforms baselines on challenging datasets. The paper discusses the limitations of the current approach, such as the need for external amplification ratios and the assumption of dedicated network training for specific camera sensors. Future work is suggested to address these limitations, including improving generalization across different sensors and optimizing runtime performance.The paper addresses the challenge of imaging in extremely low light conditions, where traditional camera processing pipelines fail due to low photon counts and low signal-to-noise ratio (SNR). The authors introduce a new dataset, the See-in-the-Dark (SID) dataset, which contains 5094 raw short-exposure images paired with corresponding long-exposure reference images. This dataset is designed to support the development of data-driven approaches for low-light image processing. To address the limitations of traditional methods, the authors propose a pipeline based on end-to-end training of a fully-convolutional network (FCN). The FCN processes raw sensor data directly, bypassing the traditional image processing pipeline, which often performs poorly in such conditions. The network is trained to perform tasks such as color transformations, demosaicing, noise reduction, and image enhancement. The SID dataset and the proposed pipeline are evaluated through various experiments, including qualitative comparisons with traditional pipelines, denoising algorithms, and burst processing. The results show that the proposed pipeline significantly outperforms existing methods in terms of noise suppression and color transformation. The authors also conduct perceptual experiments using Amazon Mechanical Turk to assess user preference, finding that their pipeline significantly outperforms baselines on challenging datasets. The paper discusses the limitations of the current approach, such as the need for external amplification ratios and the assumption of dedicated network training for specific camera sensors. Future work is suggested to address these limitations, including improving generalization across different sensors and optimizing runtime performance.