ActionDiffusion: An Action-aware Diffusion Model for Procedure Planning in Instructional Videos

Lei Shi,Paul Bürkner,Andreas Bulling

2024-07-21

Abstract:We present ActionDiffusion -- a novel diffusion model for procedure planning in instructional videos that is the first to take temporal inter-dependencies between actions into account in a diffusion model for procedure planning. This approach is in stark contrast to existing methods that fail to exploit the rich information content available in the particular order in which actions are performed. Our method unifies the learning of temporal dependencies between actions and denoising of the action plan in the diffusion process by projecting the action information into the noise space. This is achieved 1) by adding action embeddings in the noise masks in the noise-adding phase and 2) by introducing an attention mechanism in the noise prediction network to learn the correlations between different action steps. We report extensive experiments on three instructional video benchmark datasets (CrossTask, Coin, and NIV) and show that our method outperforms previous state-of-the-art methods on all metrics on CrossTask and NIV and all metrics except accuracy on Coin dataset. We show that by adding action embeddings into the noise mask the diffusion model can better learn action temporal dependencies and increase the performances on procedure planning.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the problem of procedure planning in instructional videos. Specifically, existing methods fail to fully utilize the temporal dependencies between actions when handling action sequences. The paper proposes a new diffusion model—ActionDiffusion, which is the first method to consider action temporal dependencies during the diffusion process. By embedding actions into a noise mask and introducing an attention mechanism in the denoising network, ActionDiffusion can better learn the temporal order dependencies between actions, thereby improving the performance of procedure planning. Experimental results show that ActionDiffusion outperforms existing methods on multiple benchmark datasets.