PAR Contest 2023 is a competition, organized within CAIP 2023 conference, among methods based on multi-task learning, aimed at the recognition of binary and multi-class pedestrian attributes from images. This topic is recently attracting a great interest of various research groups due to the variety of applications in the field of forensics, digital signage, social robotics, people tracking and multi-camera person re-identification. Multi-task learning allows to solve the multi-class recognition problem with a single multi-task neural network, with a learning procedure that exploits the interdependencies between different tasks to produce an efficient and effective model. To this aim, we make available for the participants the MIVIA PAR Dataset, consisting of 105,244 pedestrian images, already divided in training and validation sets, partially annotated with 5 attributes: upper clothes and lower clothes color, gender, bag, hat. The submitted methods will be evaluated in terms of mean accuracy over a private test set, including more than 20,000 images without overlaps in terms of subjects and scenarios with respect to training and validation sets. The baseline results, reported in this paper, demonstrate that the contest is challenging and that by participating to the competition it is possible to advance the state of the art in pedestrian attributes recognition.
PAR Contest 2023: Pedestrian Attributes Recognition with Multi-task Learning
Greco A.;
2023-01-01
Abstract
PAR Contest 2023 is a competition, organized within CAIP 2023 conference, among methods based on multi-task learning, aimed at the recognition of binary and multi-class pedestrian attributes from images. This topic is recently attracting a great interest of various research groups due to the variety of applications in the field of forensics, digital signage, social robotics, people tracking and multi-camera person re-identification. Multi-task learning allows to solve the multi-class recognition problem with a single multi-task neural network, with a learning procedure that exploits the interdependencies between different tasks to produce an efficient and effective model. To this aim, we make available for the participants the MIVIA PAR Dataset, consisting of 105,244 pedestrian images, already divided in training and validation sets, partially annotated with 5 attributes: upper clothes and lower clothes color, gender, bag, hat. The submitted methods will be evaluated in terms of mean accuracy over a private test set, including more than 20,000 images without overlaps in terms of subjects and scenarios with respect to training and validation sets. The baseline results, reported in this paper, demonstrate that the contest is challenging and that by participating to the competition it is possible to advance the state of the art in pedestrian attributes recognition.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
