/DBR-TAD

MIT LicenseMIT

DBR-TAD: Diffusion-Based Boundary Refinement for Temporal Action Detection

We provide the IAL model code as supplementary material and publish the DiffBR model code after acceptance of the paper.

Overview

Overview

Installation

  • Recommended Environment: python 3.8.8, Cuda11.6, PyTorch 1.12.1(The PyTorch version should be at least >= 1.11.)
  • Install dependencies: pip install -r requirements.txt
  • Install NMS: cd ./libs/utils; python setup.py install --user; cd ../..

Data

Download Features and Annotations

Dataset Feature Encoder Link
THUMOS14 I3D thumos_i3d
THUMOS14 VideoMAE thumos_videomae
ActivityNet I3D anet_i3d
ActivityNet TSP anet_tsp
ActivityNet VideoMAE anet_videomae
Epic-Kitchen SlowFast epic_kitchen

Unpack Features and Annotations

  • Unpack the file under ./data
  • The folder structure should look like
DAD-TAD/
  ├── data
  │   ├── anet_1.3
  │   │   ├── annotations
  │   │   ├── i3d_features
  │   │   ├── tsp_features
  │   │   └── anet_mae_hugek700
  │   ├── epic_kitchens
  │   │   ├── annotations
  │   │   ├── features
  │   └── thumos
  │       ├── annotations
  │       ├── i3d_features
  │       ├── th14_mae_g_16_4
  ├── libs
  ├── tools
  └── ...

Training and Evaluation

  • We have provided a script list that allows you to replicate our results with just a single click. Further details can be found in ./tools/run_all_exps.sh.

  • Our experiments were conducted exclusively on a single NVIDIA GeForce GTX 1080 Ti. It is noted that variations in the graphics card model may lead to slight discrepancies in replicating the results.

  • [Optional] Monitor the training using TensorBoard. Example:

tensorboard --logdir=./ckpt/thumos_i3d_final/logs