Raw smartphone images captured across times of day, weather conditions and light sources, with contextual metadata (time, location, capture settings), neutral and user-preferred white-balance ground truth, and expert-rendered sRGB images.
A reprocessed version of the Sony and Nikon splits of the LSMI dataset, rendered with five white-balance presets (2850 K to 7500 K) and paired with corrected ground truth, for training and evaluating multi-illuminant white-balance correction.
Paired RGB and multispectral (MS) mosaiced images captured with a dual-camera setup, with ground-truth demosaiced outputs, for training and evaluating RGB-guided multispectral demosaicing.
RAW images of hand-crafted scenes full of high-frequency detail that is difficult to demosaic, with annotated patches for training, validation and testing joint demosaicing and denoising on Single-, Quad- and Nona-Bayer sensors.
A 4D dataset that records the complete range of possible exposures over time for challenging dynamic scenes, with a software platform for repeatable evaluation of autoexposure algorithms.
Test images of scenes lit by multiple light sources, for evaluating white-balance correction in mixed-illuminant conditions.
High-quality dual-pixel captures for learning to synthesize the two dual-pixel sub-aperture views from a single image, used to improve single-image defocus deblurring through multi-task learning.
Raw image pairs captured at the same time with the main and wide-angle cameras, with colour-chart illuminant ground truth, for illuminant estimation that exploits multiple smartphone cameras.
Images rendered from the MIT-Adobe FiveK raw images, pairing under- and overexposed versions with expert-retouched references.
Each scene has a defocused wide-aperture image, its two dual-pixel sub-aperture views, and the corresponding all-in-focus image. A widely used benchmark for dual-pixel and single-image defocus deblurring.
Camera-rendered images with incorrect and correct white balance, for correcting and editing white balance in sRGB images. Also the basis for Deep White-Balance Editing (CVPR 2020).
Temporal focal-stack sequences captured on a smartphone, with a software platform that emulates phase-detection and contrast-detection autofocus for repeatable evaluation of autofocus algorithms.
A standard benchmark for developing and evaluating real-image denoising. Basis of the NTIRE 2019 and 2020 Real Image Denoising Challenges at CVPR.
Each image contains a colour chart and is paired with a colorimetric reference in linear ProPhoto RGB, for evaluating how accurately cameras reproduce colour.
Aligned images of the same scenes from different cameras, with colour-chart calibration images, for studying how raw colour responses differ between sensors and how to map between them.
Curated and labelled images for colour constancy, later extended with a laboratory set and full colour-chart ground truth in collaboration with Adobe Research. Widely used for training and evaluating colour constancy methods.
A lightweight, open-source raw-to-sRGB pipeline (white balance, demosaicing, colour space conversion, tone mapping), used to render the SIDD images and support the NTIRE denoising challenges.
Emulates camera hardware so that each stage of the in-camera processing pipeline can be studied and modified independently. Developed with support from a Google Research Award.