LearnLFOdo Dataset

Description:

This is the dataset associated with the paper "Unsupervised Learning of Depth Estimation and Visual Odometry for Sparse Light Field Cameras", S. Tejaswi Digumarti, Joseph Daniel, Ahalya Ravendran, Donald G. Dansereau.

This data was collected using an EPIModule from EPIImaging, LLC. mounted on a robotic arm while executing several trajectories.

Core dataset:

Sequences:

The core of the dataset comprises of folders named seq##. Each of these folders corresponds to one trajectory executed by the robot arm, with ## being the ID of the sequence.

Each seq## folder contains 17 subfolders 0, 1, 2, ..., 16 where the number corresponds to the id of the sub-aperture of the EPIModule.

The sub-apertures are in a plus configuration as shown below, with the camera pointing into the screen and the reader looking at the imaging module from behind.

    
            0  
            1  
            2  
            3
4  5  6  7  8  9  10  11  12
            13
            14 
            15 
            16

The sequences used for training are mentioned in train.txt
The sequences used for validation are mentioned in val.txt
Sequences 80, 81, 82 are test trajectories.

Poses

All the seq# folders also contain groundtruth poses of the robotic arm. These are organized as follows.

• poses_gt_base_endeff.npy - pose of the endeffector wrt base
• poses_gt_first_endeff.npy - pose of the endeffector wrt the first pose
• poses_gt_prev_curr_endeff.npy - pose of the endeffector wrt its previous pose
• poses_gt_base_cam.npy - pose of the camera wrt base. The x-axis of the camera frame is to the right, y-axis downwards and z-axis along the viewing direction fo the camera.
• poses_gt_first_cam.npy - pose of the camera wrt the first pose
• poses_gt_prev_curr_cam.npy - pose of the camera wrt its previous pose

Static Images for Depth Estimation

The folder depth_planes_textured was used to quantitatively evaluate depth estimation and contains images of a planar object taken at distances of 40, 50, 60, 70 and 80cm from the imaging module.

The folder depth_teaser was used to generate the teaser image in the paper.

Extra data:

Additional data not used in the paper is provided in the extras folder. This contains a few more sequences, images of another planar object without a strong texture and images of a few more objects.