GM-ATCI Rear-view pedestrians dataset





Description

This dataset was collected using a vehicle-mounted standard automotive rear-view display camera for evaluating rear-view pedestrian detection. The production 180 degrees fisheye-lens camera was mounted in the front of the vehicle for safety reasons but in a typical rear-view installation pose: 107cm height and 25 degrees downward tilt angle. The dataset contains 15 filming sessions, each taken in a different day with different scenarios. Each session contains multiple clips with duration ranging from several seconds to several minutes. In total, the dataset contains 250 clips with a total duration of 76 minutes and over 200K annotated pedestrian bounding boxes. Out of this dataset we release for download here the test portion comprising of 70 of the clips as described in [1]. There are two types of sessions, containing either staged or “in-the-wild” pedestrians. The staged scenarios include mainly pedestrians walking in front of the camera at different positions and directions in a controlled manner, spanning the different use cases of a rear alert or braking automotive feature. In the remaining sessions the vehicle drove either in public roads or in parking lots and captured incidental pedestrians. The different locations include: indoor parking lots, outdoor paved/sand parking lots, city roads and private driveways. We filmed both day and night scenarios, with different weather and lighting conditions. More information can be found in the rear-view detection system papers [1, 2]. If you make use of the dataset in your research please cite [2].

Download

GM-ATCI Pedestrian test set - consists of 6 sets of clips. The videos are saved in the Seq format defined in Piotr's matlab toolbox

You can also download an example video (Avi, Seq) from the training set along with its corresponding manual annotation (vbb).

Camera calibration: here you'll find a set of checkerboard images taken with the camera, and a calibration result we obtained using the Caltech Calibration Toolbox. Notice some of the checkerboards are on the ground and can be used to estimate the camera to ground pose.

Evaluation and results

Detection performance is evaluated using the evaluation code provided in the Caltech Pedestrian detection benchmark. We compare the bounding boxes to manual annotation, and ignore pedestrians outside the main central area (as marked by the blue dotted line in the images above). More details on the evaluation are found in [1, 2]. The results as published in [1] contain several versions of our system compared against a HoG based detector.


Submitting Results

Please contact us to include evaluation of your detector's performance on this dataset. Results should be submitted in the same format as in the Caltech Pedestrian detection benchmark

Contact

Please contact Dan Levi (dan.levi@gm.com) for any question, remark or to submit results.

References

[1] Tracking and motion cues for rear-view pedestrian detectionDan Levi and Shai Silberstein IEEE Intelligent Transporation Systems Conference (ITSC) 2015. PDF

[2] Vision-Based Pedestrian Detection for Rear-View CamerasShai Silberstein, Dan Levi, Victoria Kogan and Ran GazitIEEE Intelligent Vehicles Symposium 2014. PDF