This page provides an overview of how to use our datasets. We provide the following (see the section below for sub-directory organization):

• Example iPython Notebook for processing the original measurements for homography warping and for processing the measurements for machine learning training.
• Experimentally captured point spread function (PSF) measurements. [LINK]
• The original, full resolution RML, Diffusercam, and Ground Truth Lensed measurements. [LINK]
• Ground truth lensed measurements at 8x downsampling with lens distortion removed and aligned to lensless imagers. [LINK]
• Homographies mapping from Ground Truth to RML and Ground Truth to Diffusercam at 8x downsampling. [LINK]

Our dataset directory is organized as follows:

            100k Dataset/
            ├── (1) Full Resolution Measurements/
            │   ├── 0-25k/
            |   │   ├── diffusercam/
            |   │   ├── ground_truth/
            |   │   ├── rml/
            |   │   └── undistorted_ground_truth_images/
            │   ├── 25k-50k/
            │   ├── 50k-75k/
            │   └── 75k-100k/
            ├── (2) PSFs/
            │   ├── mono8_psfs/
            │   ├── mono12_psfs/
            │   └── rgb_psfs/
            ├── (3) Homographies/
            │   ├── Ground Truth to Lensless Imager/
            ├───└── Lensless Imager to Ground Truth/
            └───(4) 8x Downsampled Registered Ground Truth to Imager Space/
          

Folder (1) contains full resolution measurements of 100,000 images, split into 25,000 folders which each contain a folder for each imager: the Diffusercam, ground truth lensed camera, and RML. We also include the undistorted version of the lensed images.

Folder (2) contains PSFs for each imager in Mono8, Mono12, and RGB8 color formats.

Folder (3) contains .npy files storing homography transformations from the lensed camera to both lensless imagers and from the lensless imagers to the lensed camera at 8x downsampling.

Folder (4) contains 8x downsampled versions of the undistorted lensed measurements warped to the space of the lensless imagers.

25,000 Image Dataset: We also include a separate folder that follows a similar sub-directory structure for those who only want to interface with the 25,000 image dataset, linked above.

With our dataset, we provide two types of homography transforms that can be used to ensure the alignment of lensed ground truth and lensless measurements:

• Homography from Ground Truth to Lensless Imager: warps the ground truth lensed camera to the spaces of each lensless imager (Diffusercam and RML).
• Homography from Lensless Imager to Ground Truth: warps each lensless imager to the space of the ground truth lensed camera. Useful for fair comparison across all three imagers.

These are for 8x downsampled versions of the dataset. We anticipate providing homographies for 4x and 2x downsampling in the future.

We provide an iPython Notebook tutorial that walks users step-by-step through how to prepare our dataset to be used in their own projects. Using the 8x downsampled measurements as a starting point, the tutorial includes loading lensed measurements, downsampling lensless measurements, and preparing the data for training and evaluation view.