This paper introduces a new method for dataset condensation in action segmentation, moving from optimization-based inversion to deterministic latent mapping. The approach utilizes Denoising Diffusion Implicit Models to represent action segments as continuous trajectories. An adaptive allocation mechanism dynamically adjusts the anchoring budget based on reconstruction difficulty, outperforming existing methods and achieving performance parity with real data training at a 2.4% condensation ratio on the Breakfast dataset. AI
IMPACT This research could lead to more efficient training of action segmentation models by reducing data requirements.
RANK_REASON Academic paper detailing a new method for dataset condensation in computer vision.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →