Neural Imaging Resources
This page contains additional resources (images, pre-trained models, etc.) for my neural imaging toolbox.
TLDR:
training-data-D90.zip via Google drive (4.7 GB)
rgb-testsets-512px.zip via Google drive (43 MB)
Legacy Models (TF1):
Pre-trained Camera Models
The toolbox provides implementations of both neural and classing imaging pipelines. The neural pipelines (UNet, DNet and INet in models.pipelines) replace the entire ISP with a CNN and need to be trained for each camera model. The classic pipeline (ClassicISP) is universal and can be reconfigured for various cameras by setting the desired CFA configuration and sRGB color conversion tables. The classic pipeline follows the steps of standard ISPs but uses neural demosaicing. The model was pre-trained and works for various CFA configurations.
Pre-trained models can be downloaded here: models_nip_isp_dcn_tf2.zip via Google drive (59 MB). The archive contains neural pipelines for Nikon D90 and a universal classic ISP with neural demosaicing.
Pre-trained Compression Models
The toolbox comes with a reference implementation of a simple learned image compression codec (see models.compression for model definition and compression.codec for bistream encoding and decoding functions). The rate distortion performance of the included models is shown below. A standalone deployed version of the codec is available at github.com/pkorus/l3ic.
RAW-RGB Images for Camera ISP Training
Training data for Nikon D90 can be downloaded here: training-data-D90.zip via Google drive (4.7 GB). The archive contains 150 pairs of RAW + RGB full-resolution images (selected randomly from the Raise dataset). The RAW images are provided as numpy arrays arranged as 4-channel RGGB Bayer stacks with applied camera white-balance. The RGB images were developed using a simple camera ISP developed in Python (see helpers.raw_api).
RGB Images
We tesed our compression codec on 3 different datasets: (clic) 39 validation images from the 2019 CLIC compression challenge; (kodak) 24 standard Kodak test images; and (raw) 39 patches with native camera output (no post-processing - e.g., resizing - after the camera ISP). All images are uncompressed 512 x 512 px. RGB bitmaps. The dataset can be downloaded here: rgb-testsets-512px.zip via Google drive (43 MB). The archive also contains rate-distortion profiles for JPEG, JPEG 2000, BPG and DCN codecs.
