Surrey Mobile Dataset

Copyright Notice

University of Surrey is the copyright holder of all the images included in the dataset. If you use this dataset, please cite the following paper:
Simone Madeo and Miroslaw Bober: "Fast, Compact and Discriminative: Evaluation of Binary Descriptors for Mobile Applications".
This dataset is provided as is and without any express or implied warranties, including, without limitation, the implied warranties of merchantability and fitness for a particular purpose.

Dataset Description

The dataset consists of 8658 frames extracted from video sequences for evaluation of image search on mobile scenario. The sequences show a table containing several objects, including postcards, little boxes, papers, flowers, glasses, bottles, business cards and magazines. Three different devices have been used: GoPro, Nexus 7 and iPhone. Videos have been recorded in different times of the day, with natural and artificial light generating different shadow shapes and reflections. Videos are also subject to blurring and different zooms of the target object. The objects have been setup in order to create occlusions or transparency effects. Moreover, several features on different queries could match the same features on a given database image, producing false positive matches.

The dataset can be downloaded from this page, see details below. The material given includes:

  • the images, divided into 3 folders depending on the device used,
  • 64x64 image patches representing Hessian-Affine regions,
  • the evaluation software.

Download and Statistics

Dataset size: 8658 video frames in pgm format (1920 x 1080) extracted at 5 fps, including 17 sequences for GoPro, 11 sequences from iPhone, 7 sequences from Nexus7.
Queries: 6 image queries in pgm format.
Download the video frames (tar.gz, 15 GB)
Download the image patches (tar.gz, 72 GB)
Download the evaluation software (tar.gz, 1 MB)

Evaluation Protocol

In order to assess the presence of a query object in a frame, for each frame and for each object a score can be computed by means of a custom matching algorithm. By thresholding the score values, a binary map can be built and compared to the ground-truth. The result is the computation of true positive ratio (TPR) and false positive ratio (FPR) metrics.
Please read the documentation for more details.