|Description (include details on usage, files and paper references)||Daimler Stereo Pedestrian Detection Benchmark
C. Keller, M. Enzweiler, and D. M. Gavrila, A New Benchmark for Stereo-based Pedestrian Detection, Proc. of the IEEE Intelligent Vehicles Symposium, Baden-Baden, Germany, 2011.
This new benchmark extends the previously published Daimler Mono Pedestrian Detection Benchmark with
- 7129 stereo image pairs not containing pedestrians in the training set, from which negative samples can be extracted (positive samples remain unchanged)
- images of the second camera for the test set (a sequence with more than 21.790 images with 56.492 pedestrian labels, fully visible or partially occluded, captured from a vehicle during a 27 min drive through urban traffic, at VGA resolution (640x480, uncompressed)).
- vehicle speed and steering angle measurements for the test set, with the yaw rate derived.
- specifies an evaluation setting (3D localization criterion, known ground plane, and sensor coverage area provides ROIs).
- specifies performance metrics both at the frame- and trajectory-level (the latter also allows benchmarking of tracking algorithms).
- provides the baseline performance of a state-of-the-art method (HOG/linSVM) on the specified training and test set.
- is open: both training and test set are public and (freely) available for non-commercial purposes, see below.