SEAL-Pose: Enhancing Pose Estimation through Trainable Loss Function

Junggeun Do, Jay-Yoon Lee

Keywords: Structured Energy network, Energy-based models, Trainable Loss-function, Dynamic loss function, Pose Estimation

Abstract:

Accurately predicting 3D human pose is a challenging task in computer vision due to the need to capture complex spatial structures and anatomical constraints. We propose SEAL-Pose, an adaptation of the Structured Energy As Loss (SEAL) framework for deterministic models, specifically designed to enhance 3D human pose estimation from 2D keypoints.
Although the original SEAL was limited to probabilistic models, our approach employs the model's predictions as negative examples to train a structured energy network, which functions as a dynamic and trainable loss function.
Our approach enables a pose estimation model to learn joint dependencies via learning signals from a structured energy network that automatically captures body structure during training without explicit prior structural knowledge, resulting in more accurate and plausible 3D poses .
We introduce new evaluation metrics to assess the structural consistency of predicted poses, demonstrating that SEAL-Pose produces more realistic, anatomically plausible results.
Experimental results on the Human3.6M and Human3.6M WholeBody datasets show that SEAL-Pose not only reduces pose estimation errors such as Mean Per Joint Position Error (MPJPE) but also outperforms existing baselines.
This work highlights the potential of applying structured energy networks to tasks requiring complex output structures, offering a promising direction for future research.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 14045