Challenge on Thermal Human Pose Estimation

Important dates

Registration Open: June 13, 2024

Training Data Online: June 13, 2024

Test Data Release: July 13, 2024

Challenge Submission Deadline: July 31, 2024

Results Announcement: August 7, 2024

Challenge Reports Deadline: August 22, 2024

Camera-Ready Papers Due: September 10, 2024

ICPR 2024 dates: December 01-05, 2024


RGB-Thermal Nearly Paired and Annotated 2D Pose Dataset

We are excited to introduce a novel dataset that addresses the challenges of pose estimation in Long-Wave Infrared (LWIR) thermal images. Our dataset, comprising over 2,400 high-quality LWIR images, is meticulously annotated with 2D human poses, providing a valuable resource for researchers and practitioners. The dataset is designed to facilitate pose estimation in occlusion and other challenging scenarios, making it an ideal benchmark for evaluating the performance of pose estimation models in real-world applications.

Dataset Characteristics

Our dataset is unique in that it introduces deliberate complexities that reflect real-world scenarios. The dataset encompasses a diverse range of:

  • Pose variations: from simple standing positions to complex actions like walking and greeting
  • Body shapes: including subjects with different body types
  • Clothing: diverse clothing styles to encourage pose estimation independent of skin texture
  • Self-occlusion: poses with self-occlusion to challenge the model to infer pose even when parts of the body are hidden from view
  • Different activities: various activities like sitting, walking, and exercise to expose the model to a broader range of motion patterns

By incorporating these complexities, the dataset aims to train models that are robust to the challenges inherent in IR pose estimation and can perform accurately in real-world applications.

Dataset Details

For more information about the dataset, please refer to our paper:

To download the dataset and take part in the competition please, fill the registration form.

Evaluation Metric

The performance of the models will be evaluated using two metrics:

  1. Mean Per Joint Position Error (MPJPE): This metric measures the average Euclidean distance between the ground truth 2D pose keypoints and the estimated 2D pose keypoints. The MPJPE loss is calculated as:

L_res = (1/n) * ∑[k=1 to n] ||p_k,i – p̂_k,i||²

where n is the number of joints in a pose, p_k,i is the ground truth 2D pose keypoint, and p̂_k,i is the estimated 2D pose keypoint. A lower MPJPE value indicates better model performance.

  1. Percentage of Correct Keypoints (PCKh): This metric measures the accuracy of the localization of different critical points with a given threshold. Specifically, PCKh@0.5 will be used, which is 50% of the head bone link. A higher PCKh value represents better model performance.

Participation Link

Registration will be done through the below link

Registration Form

Please upload signed EULA form on the Google form link as given above. Click on the below link to download the EULA form:


Once you complete the registration, you will receive a mail from the organizers containing link to download the dataset. If you do not receive the dataset link and participation confirmation, kindly write a mail to or

Challenge Rules

The challenge is open for participants from industry and academia, who wants to make a contribution to the field of Infrared Computer Vision. Below we outline the challenge rules.

  • Each participant team can include up to a maximum of 5 people from one or more affiliations. For the sake of fairness to smaller research groups, we will not allow bigger teams to participate as a single team.
  • One person can only participate on one team. Mentors included. No exceptions.
  • Winners teams will have certificates listing the names of their members in the exact format and order as they were registered.
  • We will only allow registrations to be updated until the date of the release of the testing dataset. Changes allowed before this date include adding or removing members, correcting typos on names, as well as updating the order of the team members which will be used for the certificate in the case of winners.
  • Participants can register to participate in individual tasks only. However, winners will be selected based on their task coverage. We will be using a scoring system.
  • Participants will be allowed to use only datasets provided for training.
  • Submissions will be done as per the instructions given on the website.
  • CODE submission will be required.
  • Participants must provide a description of the methods used to produce the results submitted. In the final competition paper, we will summarize these descriptions when we describe the submitted systems. We reserve the right to disqualify submissions that do not provide a sufficiently detailed description of their system.
  • To be fair to all participants, any deadline extensions given will apply to all participants, not just to individual research groups who might request them.