Novelty Unlocking with Multiobjective Generative Models: Batch Diversity of Human Motions
Keywords: Multiobjective optimization;Diverse In-Betweening Human Motions
TL;DR: Unlocking Novelty through Multiobjective Generative Models: A Study on Diverse In-Betweening Human Motions within Batch Diversity
Abstract: Current generative models have shown potential performance in many tasks, which typically focus on generating samples that closely adhere to a given distribution, often overlooking the requirement to produce optimal diverse solutions in a batch diversity.
Recognizing that maintaining ``diversity" has been a longstanding challenge in multiobjective optimization, we were inspired to introduce a multiobjective optimization approach to enhance diversity in a single pass.
This paper utilizes the in-betweening human motion generation task as an example and introduces the multiobjective generative models to demonstrate the effectiveness of the proposed method in producing diverse and smooth human motion sequences. The resulting method, termed the \textit{Multiobjective Generation Framework with In-Betweening Motion Model} (MGF-IMM), frames the human motion in-betweening task as a bi-objective optimization problem. The designed in-betweening motion model is then integrated into a nondominated sorting-based optimization framework to address this bi-objective optimization problem.
Through comprehensive qualitative and quantitative experiments, MGF-IMM has demonstrated state-of-the-art performance, surpassing the latest methods and validating its superiority in generating diverse in-betweening human motions.
Primary Area: generative models
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: 4357
Loading