DiffGEBD: Generic Event Boundary Detection via Denoising Diffusion (ICCV 2025)

Official Implementation of our ICCV Paper https://arxiv.org/pdf/2508.12084:

DiffGEBD: Generic Event Boundary Detection via Denoising Diffusion
Jaejun Hwang*, Dayoung Gong*, Manjin Kim, and Minsu Cho
ICCV, Hawaii, 2025

---

1. Installation

System Requirements

• Python 3.10.13
• CUDA 11.8
• PyTorch 2.1.0

Install Environments

conda env create -f requirements.yaml
conda activate base

2. Dataset Setup

Kinetics-GEBD

Download Kinetics-400 from https://github.com/cvdfoundation/kinetics-dataset and uniformly sample 100 frames. The image files should be named as 'img_*****.jpg'.

*Some videos cannot download from YouTube. ./data/file_list.pkl contains names of the available videos. If any video in the list is unavailable, modify the file_list.pkl accordingly.

TAPOS

Download TAPOS from https://opendatalab.com/OpenDataLab/TAPOS/download or https://sdolivia.github.io/TAPOS/.

3. Training

Set your dataset path and start training:

CUDA_VISIBLE_DEVICES=<GPU_IDS> \
        torchrun --nproc_per_node=<NUM_GPUS> --master_port=<PORT> train.py --local_rank 0 \
            --config-file ./config/config.yaml \
            --gebd_data_dir /path/to/your/kinetics_folder/ \
            --tapos_data_dir /path/to/your/tapos_folder/

Note:

• Replace <GPU_IDS> with your GPU IDs (e.g., 0,1,2,3 for 4 GPUs)
• Replace <NUM_GPUS> with the number of GPUs you want to use
• Replace <PORT> with an available port number (e.g., 10210)
• Modify --gebd_data_dir and --tapos_data_dir to point to your dataset directories
• You can also use environment variables: export GEBD_ROOT=/path/to/kinetics and export TAPOS_ROOT=/path/to/tapos

4. Inference

1. Download checkpoints and config files from this link

2. Set the paths and run inference:

CUDA_VISIBLE_DEVICES=<GPU_IDS> \
        torchrun --nproc_per_node=<NUM_GPUS> --master_port=<PORT> train.py --local_rank 0 \
            --config-file ./checkpoints/kinetics-gebd/config.yaml \
            --gebd_data_dir /path/to/your/kinetics_folder/ \
            --test-only \
            --resume ./checkpoints/kinetics-gebd/model_best.pth \
            --seed 42 \
            SOLVER.BATCH_SIZE 2 \
            DIFFUSION.SAMPLING_TIMESTEPS 32 \
            DIFFUSION.CFG_SCALE 7.0

Note:

• Replace <GPU_IDS> with your GPU IDs (e.g., 0,1,2,3 for 4 GPUs)
• Replace <NUM_GPUS> with the number of GPUs you want to use
• Replace <PORT> with an available port number (e.g., 10210)
• Modify --config-file and --resume paths to your downloaded checkpoint and config files
• Adjust --gebd_data_dir to your dataset path
• You can modify SOLVER.BATCH_SIZE, DIFFUSION.SAMPLING_TIMESTEPS, and DIFFUSION.CFG_SCALE as needed

5. Calculating Metrics

Use jupyter-notebook ./utils/calc_metrics.ipynb

• For Kinetics-GEBD, ground-truth file is ./data/k400_mr345_val_min_change_duration0.3.pkl
• For TAPOS, ground-truth file is ./data/TAPOS_for_GEBD_val.pkl

6. Acknowledgement & Citation

This repository builds upon the EfficientGEBD codebase. We thank the original authors for sharing their work.

If you find our code or paper helpful, please consider citing both DiffGEBD and EfficientGEBD:

@inproceedings{hwang2025generic,
  title={Generic Event Boundary Detection via Denoising Diffusion},
  author={Hwang, Jaejun and Gong, Dayoung and Kim, Manjin and Cho, Minsu},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={14084--14094},
  year={2025}
}

@inproceedings{zheng2024rethinking,
  title={Rethinking the architecture design for efficient generic event boundary detection},
  author={Zheng, Ziwei and Zhang, Zechuan and Wang, Yulin and Song, Shiji and Huang, Gao and Yang, Le},
  booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
  pages={1215--1224},
  year={2024}
}

---

📄 License

This project is licensed under the MIT License.