3D Human Pose Estimation Using Ultra-low-resolution Thermal Images

Can we perceive the three-dimensional posture of the whole human body solely from extremely low-resolution thermal images (e.g., $8\times8$-pixels)?

This paper investigates the possibility of this challenging task.

Thermal images capture only the intensity of radiation, making them less likely to contain personal information such as facial or clothing features.

Thermal sensors are commonly integrated into daily-use appliances, such as air conditioners, automatic doors, and elevators.

Considering these advantages, we propose a framework for estimating three-dimensional (3D) human pose by utilizing very low resolution (only 8-pixel square) thermal images. This not only reduces the inclusion of personal information, but also opens up memory-efficient measurement capabilities with potential applications in various fields.

Compared to conventional cameras, thermal images are more susceptible to variations in room temperature and subject-specific differences, such as physical build or basal body temperature.

To address this challenge, we apply attribute-independent feature extraction. Specifically, we employ discriminators designed to perform feature extraction independent of subject and ambient temperature, ensuring that the extracted features are invariant to these attributes. This approach enhances the robustness of our method by enabling it to focus on the core posture-related features rather than variations due to subject or ambient temperature differences.