Data

New College	Format	Size
Images	ZIP	546 MB
GPS (Matlab Format)	Binary	28 KB
GPS (Plain Text)	Text	69 KB
Ground Truth Mask (Matlab Format)	Binary	21 KB
Ground Truth Mask (Plain Text)	ZIP	18 KB
Aerial Photo	JPEG	400 KB

City Centre	Format	Size
Images	ZIP	546 MB
GPS (Matlab Format)	Binary	28 KB
GPS (Plain Text)	Text	69 KB
Ground Truth Mask (Matlab Format)	Binary	27 KB
Ground Truth Mask (Plain Text)	ZIP	21 KB
Aerial Photo	JPEG	1 MB

Description

Each dataset includes the following components:

Images

Images are numbered sequentially in order of collection. The camera was mounted on a pan-tilt, and collects images to the left and right of the robot. Image collection was triggered every 1.5m (on the basis of odometry). The vehicle is in motion while the images are collected, so the robot travels some distance between the collection of the right and left images.

Image Collection Coordinates (GPS)

(X,Y) coordinates are supplied for each image, corresponding to robot position at the time the image was acquired. The coordinates are derived from interpolated GPS. Where GPS signal was unavailable (substantial sections of the New College dataset), fake GPS messages were supplied by hand. As such, the position data is approximate and is intended for visualization purposes only.

Ground Truth

Algorithm performance is measured in terms of the image-to-image correspondence matrix C generated by the loop closing algorithm. Entry (i,j) of the correspondence matrix is 1 if image i and image j were determined to be taken at the same place, and 0 otherwise.

Ground truth is specified as a "mask" for this correspondence matrix. The mask is a binary matrix M the same size as the correspondence matrix, which identifies most false positives. Correspondences identified by the algorithm which lie in a zero-region of the mask are marked as false positives. Due to the labour involved in hand-labeling all image-to-image matches, the masks are approximate, and true and false positives labeled in this way need to be visually inspected to confirm that they are correct.

Entry (i,j) of the mask matrix is 0 if image i and image j were taken at the different places.

When entry (i,j) of the mask matrix is 1, image i and image j are accepted as a potential true positive, subject to visual inspection.

Aerial Photo

Aerial photos of the data collection region are included for the purpose of visualizing results.