|
Home
Dataset/Download
Experiments
About
Contact:
SOCIA Lab. - Soft
Computing and Image Analysis Group
Department of
Computer Science, University of Beira Interior,
6201-001 Covilhã, Portugal
hugomcp@di.ubi.pt
|
P-DESTRE
Fully Annotated
Datasets for Pedestrian Detection, Tracking, Re-Identification and Search from Aerial
Devices
News:
|
03-12-2020:
A new file is available at the
"Dataset/Download" section, containing the
cropped regions-of-interest (ROIs) of
all subjects in the P-DESTRE dataset, in
".jpg" format.
|
|
12-07-2020:
The annotation files were updated. The
repeated annotations per ID/frame and the
bounding boxes with negative coordinates
(i.e., partially out-of-screen objects) were
removed.
|
|
21-05-2020:
We updated the annotation files, and added
head pose information (yaw, pitch and roll
angles), obtained according to the Deep
Head Pose [1] method.
[1] N. Ruiz, E. Chong and J. Rehg.
Fine-Grained Head Pose Estimation Without
Keypoints. In proceedings of the IEEE
Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops, doi:
10.1109/CVPRW.2018.00281, 2018.
|
|
01-04-2020:
We are pleased to announce the availability of
the first (as per April, 2020) fully
annotated freely available dataset for
supporting the research about pedestrian
1) detection; 2) tracking; 3)
re-identification and 4) search methods
from aerial data.
|
A large number of applications using unmanned
aerial vehicle (UAV) sensors and platforms is
being developed, for agriculture, logistics,
recreational and military purposes. A branch of
these applications uses the UAV exclusively for
remote sensing purposes (RS), acquiring either
top-view or oblique data that can be further
processed at a centralized node.
Simultaneously, being at the core of video
surveillance analysis, growing research efforts have
been putted in the development of pedestrian
re-identification and search methods able to
work in real-world conditions, which is seen
as a grand challenge. In particular, the
problem of identifying pedestrians in crowded scenes
based on very low resolution and partially occluded
data becomes much harder in the
multi-camera/multi-session mode, when matching data
acquired in different places and with time lapses
that deny the use of clothing information.
To date, the evaluation of pedestrian identification
techniques has been conducted mostly on tracking
databases (such as PETS, VIPeR, ETHZ and i-LIDS),
with limited availability of soft biometric
information, or even on gait recognition datasets
(e.g., CASIA), which data acquisition conditions are
highly dissimilar of the typical occurring in
surveillance environments.
As a tool to support the research on pedestrian
detection, tracking, re-identification and search
methods, the P-DESTRE is a multi-session
dataset of videos of pedestrians in outdoor
public environments, fully annotated at the
frame level for:
1) ID. Each pedestrian has a unique
identifier that is kept among the data acquisition
sessions, which enables to use the dataset for
pedestrian re-identification and search problems;
2) Bounding box. The relative position of
each pedestrian in the scene is provided as a
bounding box, for every frame of the dataset, which
also enables to use the data for object
detection/semantic segmentation purposes;
3) Head Pose.
3D head pose information is provided in terms of
"yaw", "pitch" and "roll" angles for all the
bounding boxes (except backside views);
4) Soft biometrics. Each subject of the
dataset is fully characterised using 16 labels: gender,
age, height, body volume,
ethnicity, hair color, hairstyle,
beard, mustache, glasses, head
accessories, action, accessories
and clothing information (x3), which enables
to use the dataset also for evaluating soft
biometrics inference and
inference
techniques.
The P-DESTRE
data were acquired inside the campi of two different
universities, with students offering themselves as
volunteers of the data acquisition sessions.
1) The University of Beira
Interior, Portugal:
2) The JSS Science and
Technology University, India:
Task 1: Pedestrian
Detection
The P-DESTRE dataset
provides a bounding box for defining the region-of-interest
regarding every pedestrian in each frame of any
scene. This information is provided in the following
way:
- [x, y,
height, width]; with (x,y) being the coordinates
of the upper left corner of a pedestrian
region-of-interest with "height" and "width"
dimensions
Examples:
Task 2: Pedestrian
Tracking
Tracking information is provided implicitly by the
detection information of each pedestrian in the
scene, together with an ID that provides an unique
identifier for each subject. This enable to get
the sequences of positions of each person along
time.
Tracking Example 1:
Tracking Example 2:
Task 3:
(Short-Term) Pedestrian
Re-Identification
This is the task of associating images of the same
person taken in different occasions of the same day,
i.e., assuming that subjects keep the same
clothes between the different images.
Example 1:
Example 2:
Task 4: (Long-Term)
Pedestrian
Search
In opposition to
Re-Identification, this is a more challenging task,
and aims at the association of images of the same
person using data acquired in different days, where
no clothing information can be reliably used.
Example 1:
Example 2:
|
|
|
|
