HTTPS SSH

IRI freestyle motocross dataset

Michael Villamizar

mvillami@iri.upc.edu

http://www.iri.upc.edu/people/mvillami/

institut de robòtica i informàtica CSIC-UPC

08028 Barcelona, Spain

General Description

This dataset was created for testing detection approaches considering object rotations in the image plane. In particular, this dataset contains motorbikes under multiple orientations and with difficult imaging conditions such as partial occlusions, scale variations, lighting and intra-class changes, etc.

Training Data

For training, two sets of images are used. The first one has a collection of positive samples (object images), whereas the second one includes background or negative samples. These sets of images were extracted from the Caltech motorbikes dataset [1]. For the positive set, however, a reduced number of images (65 images) has been chosen according to the criterion of selecting the most related samples to the target motorbike model. As negative images, the Caltech background dataset is used [1]. They represent the non-object class by means of 900 example images of background. For more information about training details refer to [2].

Test Data

This dataset contains two sets of test images, the first one -test1- includes images with one or more motorbikes without rotations in the image plane, hence, they are facing right. This set contains 69 images that includes 78 motorbike instances. On the other hand, the second set of images -test2- has motorbikes with planar rotations. More specifically, this set contains 100 images with 128 motorbikes instances. Both sets were acquired from images in internet and contain challenging conditions for object detection.

File Format

For testing, one annotation file per test image is given. This file includes some relevant image information and the ground truth by means of a bounding box.

More precisely, every annotation file is provided in *.mat and contains the following information:

imgName -> image name.

numObjects -> the number of objects in the image.

objects -> an struct containing the category, bounding box -bbox-, an informative color, and the orientation of every object -angle- in the image.

In addition:

bbox -> follows the bounding box format: [left, top, width, height].

angle -> the object orientation is given in degrees.

Code

In order to show the ground truth over the test images, a matlab file -fun_bounding_box- is also provided. The execution of this file shows every test image and its correspoding bounding boxes according to the annotation file. The rest files are neccesary to execute the former file.

References

[1] R. Fergus, P. Perona, and A. Zisserman. Object class recognition by unsupervised scale-invariant learning. In CVPR, 2003.

[2] M. Villamizar, F. Moreno-Noguer, J. Andrade-Cetto and A. Sanfeliu. Efficient Rotation Invariant Object Detection using Boosted Random Ferns. In CVPR, 2010.