sauradip / diffusiontad Goto Github PK

DiffTAD is the first formulation of diffusion model for Temporal Action Detection task.

DiffTAD: Temporal Action Detection with Proposal Denoising Diffusion
Sauradip Nag, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang
arXiv 2303.14863

• (03/2023) Code will be released soon.

• First DDIM based denoising diffusion framework for Temporal Action Detection (TAD) task.
• Action Proposals are denoised in a elegant DETR style framework where Transformer-Decoder is the denoiser.
• Diffusion sampling efficiency and accuracy is enchanced by introducing a novel cross-step selective conditioning during inference.
• Denoising action proposals as queries enables faster convergence.
• First work which solves discrete dense detection task using Transformer Decoder as denoiser.

We propose a new formulation of temporal action detection (TAD) with denoising diffusion, DiffTAD in short. Taking as input random temporal proposals, it can yield action proposals accurately given an untrimmed long video. This presents a generative modeling perspective, against previous discriminative learning manners. This capability is achieved by first diffusing the ground-truth proposals to random ones (i.e., the forward/noising process) and then learning to reverse the noising process (i.e., the backward/denoising process). Concretely, we establish the denoising process in the Transformer decoder (e.g., DETR) by introducing a temporal location query design with faster convergence in training. We further propose a cross-step selective conditioning algorithm for inference acceleration. Extensive evaluations on ActivityNet and THUMOS show that our DiffTAD achieves top performance compared to previous art alternatives.

If you use DiffTAD in your research or wish to refer to the baseline results published here, please use the following BibTeX entry.

@article{nag2023difftad,
      title={DiffTAD: Temporal Action Detection with Proposal Denoising Diffusion},
      author={Sauradip Nag and Xiatian Zhu and Jiankang Deng and Yi-Zhe Song and Tao Xiang},
      journal={arXiv preprint arXiv:2303.14863},
      year={2023}
}