Download all data: data-tud.zip (2.7 MB)

Ground Truth


All annotations were done by manually placing bounding boxes around pedestrians and interpolating their trajectories between key frames. The trajectories were then smoothed to avoid discontinuous or jittery movements. All targets were annotated in all sequences, even in case of total occlusion. Each person entering the field of view acquires a unique ID, i.e. if a person leaves the screen and later reappears again, a new ID is assigned. Please note that bounding boxes are not always perfectly aligned due to articulation, interpolation and mistakes made by the annotator.
The annotations are saved in xml files using the CVML Specification, similar to the ground truth of the CAVIAR dataset.
For each dataset there are two ground truth files. One (complete) contains annotations for all visible targets, the other one (cropped) only contains targets within the predefined tracking area used in our experiments.
Please cite our work if you use the provided ground truth.

PETS 2009

All annotations were done in View001. The PETS 2009 dataset is available here.
PETS-S2L1 Ground Truth
Tracking Area in world coordinates (xmin, xmax, ymin, ymax): (-14069.6, 4981.3, -14274.0, 1733.5)

S2.L1 (12-34)

S2.L2 (14-55)

S2.L3 (14-41)

S1.L1-1 (13-57)

S1.L1-2 (13-59)

S1.L2-1 (14-06)

S1.L2-2 (14-31)

S3.MF1 (12-43)

complete
cropped
complete
cropped
complete
cropped
complete
cropped
complete
cropped
complete
cropped
complete
cropped
complete
cropped


TUD-Stadtmitte

We here provide our own annotations for the widely used TUD-Stadtmitte sequence, where bounding boxes for all targets inside the field of view (even the ones entirely occluded) are annotated. The TUD dataset is available here.
TUD-Stadtmitte Ground Truth
Tracking Area in world coordinates (xmin, xmax, ymin, ymax): (-19, 12939, -48, 10053)

TUD-Stadtmitte

complete
cropped

The camera calibration for the TUD-Stadtmitte sequence is here. It has the same format as the calibration for the PETS dataset.


TUD-Campus and TUD-Crossing

We extend the original ground truth for these two sequences by providing correct association of people identities across frames. The bounding boxes positions and sizes are kept unchanged. Additionally, we linearly interpolate the annotations through occlusions, such that, again, all targets are annotated, regardless of their visibility fraction.
TUD-Campus TUD-Crossing

TUD-Campus

TUD-Crossing

original
interpolated
original
interpolated

ETH-Person

Similar to TUD, we extend the original ground truth for several ETH sequences by providing correct association of people identities across frames. The bounding boxes positions and sizes are kept unchanged. The interpolated version for occluded pedestrians will appear shortly.
ETH-Person-seq00 ETH-Person-bahnhof ETH-Person-jelmoli ETH-Person-sunnyday

Sequence #0

Bahnhof

Jelmoli

Sunny Day

original
interpolated
original original original

Detections


To facilitate the comparison between tracking methods, we also provide the individual detections that we use for our tracking-by-detection approach.
Similar to the ground truth data, the detection bounding boxes are stored in xml files with the only difference, that the target id is missing. Instead, the detector's confidence for each bounding box is provided.

PETS 2009


S2.L1 (12-34)

S2.L2 (14-55)

S2.L3 (14-41)

S1.L1-1 (13-57)

S1.L1-2 (13-59)

S1.L2-1 (14-06)

S1.L2-2 (14-31)

S3.MF1 (12-43)

detections detections detections detections detections detections detections detections


TUD

Campus

Crossing

Stadtmitte

detections detections detections

EPFL

dataset here

terrace1

terrace2

detections c0
detections c1
detections c2
detections c3
detections c0
detections c1
detections c2
detections c3

AFL Dataset


AFL1 AFL2
The AFL Dataset accompanies our ECCV Workshop paper and contains the two video sequences, ground truth annotations and the detections used.
Download