MotionFix: Text-Driven 3D Human Motion Editing

Published: Dec. 3, 2024

The focus of this paper is 3D motion editing. Given a human and textual description the desired modification, our goal to generate an edited as described by text. challenges include lack training data design model that faithfully edits source motion. In paper, we address both these challenges. We build methodology semi-automatically collect dataset triplets in form (i) motion, (ii) target (iii) edit text, create new MotionFix dataset. Having access such allows us train conditional diffusion model, TMED, takes text input. further various baselines trained only on text-motion pairs datasets, show superior performance triplets. introduce retrieval-based metrics for editing, establish benchmark evaluation set MotionFix. Our results are encouraging, paving way research finegrained generation. Code, models available at project website.

Language: Английский

A review of deep learning-based approaches to sign language processing

Advanced Robotics, Journal Year: 2024, Volume and Issue: unknown, P. 1 - 19

Published: Dec. 23, 2024

Language: Английский