PLUS Faces in a Queue (PLUSFiaQ) - Download Page

PLUS Faces in a Queue (PLUSFiaQ)

The PLUS Faces in a Queue (PLUSFiaQ) is a multi-face tracking dataset. The dataset consists of 7 different sequences in which a total of 12 different persons move towards a gate. Annotations provided comprise a bounding box for each face, a corresponding tracking ID and a target class flag. In addition, a visibility measure and a ‘next at gate’ flag are included, but these have not been fully adjusted and corrected at this stage. Furthermore, the alignment of the bounding boxes is not defined and the contained areas of the face may differ between the bounding boxes

Table: Overview of the available sequences, where F denotes the frames , P the individual person, O_D the occlusion duration, P/F the persons per frame, O/P the ocllusions per person, μ the mean and σstandard deviation.

ID	#F	#P	μ(P/F)	μ(OD)	σ(OD)	min(OD)	max(OD)	μ(O/P)	Description
01	1774	3	0.82	52.00	32.53	29	75	0.67	simple queue
02	3051	12	3.98	44.52	55.30	1	500	22.17	eating, mask
03	701	12	3.42	42.04	50.38	1	250	3.83	pushing, eating, mask
04	576	10	3.91	38.26	38.57	2	206	4.60	queue jumping, eating
05	3126	11	3.45	39.35	41.88	3	271	19.82	stoop, walk backwards, eating, chatting, mask, fast motion
06	1301	12	3.21	43.68	39.17	4	197	7.75	walk backwards, chatting, looking down, mask
07	2201	10	3.16	43.50	45.97	1	257	15.30	looking sidewards, cover face, cover camera, mask, spoofing

Scenario

The assumed scenario is that a group of people is moving towards a gate (e.g., to enter a sports stadium). As they move towards a single gate, they form a queue. However, as this is a leisure activity, this queue will not be very well organized (people chat with each other, they eat, they move around in a disorderly fashion, there may be pushing and shoving, etc.). In the assumed scenario, the main objective is to track people's faces as they move towards the gate. For this purpose, a camera is mounted on the gate. The camera is directed at the queue such that the first person visible is standing directly in front of the gate. However, since the queue is probably not very well organized, the main tracking challenges are occlusions (partially and full), out-of-plane rotations (lateral faces) and non-linear motion. As soon as a person has entered the gate, they leave the visible area of the camera and thus the scene. To simulate multiple runs, each person goes more than once through the gate. For this reason, the people move in a counterclockwise circle.

Tracking IDs

The tracking ID is a four-digit number (e.g., xxyy), where the first two digits (xx) represent the personal identifiers (see Figure 1) and the last two digits (yy) represent a consecutive tracking number (per sequence). A new tracking number is assigned when a person has left the scene (goes through the gate) and re-enters the scene again. For all non-target class objects (faces) a personal identifier of 99 is assigned. The individual non-target class instances are assigned a dedicated tracking number, independent of how often they have re-entered the scene.

Further details can be found in:

R. Joechl and A. Uhl, "Faceqsort: a multi-face tracking method based on biometric and appearance features," arXiv preprint arXiv:2501.11741, 2025.

Filename and Directory Structure

The following directory structure applies:

• [sequnce ID], e.g. PLUSFiaQ_seq01.
  • [frames], contains all frames of the respective sequence.
• groundtruth
• annotation_files

Frame filenames are encoded according to the following structure:

• Frames: [sequnce ID]-[6-digit consecutive number].png
• Groundtruth: GT_[sequnce ID].csv
• Annotation: via_[sequnce ID].json

Annotations ChokePoint Dataset

Annotations are provided for 6 sequences (i.e., P2E_S5_C1_1, P2E_S5_C2_1, P2E_S5_C3_1, P2L_S5_C1_1, P2L_S5_C2_1 and P2L_S5_C3_1) of the ChokePoint dataset. The annotations and truth data provided have the same format as the PLUSFiaQ dataset. In the ChokePoint dataset, people are reflected in the open glass door as they walk through the portal. These reflections are annotated as non-target classes (i.e., trID 99 and the individual identifier of the reflected person).

Obtaining the Dataset

To obtain the PLUSFiaQ dataset you have to agree to our license agreement:
coming soon...

Please download, fill in and sign the license agreement and send it to R. Joechl. After checking the license agreement you will be provided with a download link.