Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



17 Commits

Repository files navigation

Diagnosing Error in Temporal Action Detectors (DETAD)

This repository is intended to host the diagnosis tool for analyzing temporal action localization algorithms. It includes three analyses: false positive analysis, localization metric sensitivity analysis, and false negative analysis. Please take a look at our DETAD paper, in which we first proposed this tool and presented an in-depth diagnosis of the state-of-the-art temporal action localization algorithms on both ActivityNet v1.3 and THUMOS14.

If you find any piece of code valuable for your research, please cite this work:

  title={Diagnosing Error in Temporal Action Detectors},
  author={Alwassel, Humam and Caba Heilbron, Fabian and Escorcia, Victor and Ghanem, Bernard},
  booktitle={The European Conference on Computer Vision (ECCV)},

How to run it?

⚠️ We have updated the code to work correctly on Python 3. Please pull the latest version and install the new conda environment + system packages as detailed below.

Install the following system packages (LaTex + fonts):

sudo apt-get install texlive-latex-extra texlive-fonts-recommended cm-super

Install conda, create the environment from the environment.yml file, and then activate the environment:

conda env create -f environment.yml
conda activate detad

Run the false positive analysis (the other two analyses are run in a similar way):

cd src
python --ground_truth_filename ../data/activity_net_train_val_extra_characteristics.v1-3.min.json --subset validation --prediction_filename ../data/anet18_winner_validation.json --output_folder ../sample_output/activitynet/validation

The code will produce PFD files with the analysis results in the --output_folder specified. Take a look at the sample_output folder for sample analysis results for a sample method on the validation and testing subset of ActivityNet v1.3 dataset and a sample method on the testing subset of THUMOS14 dataset. We thank Tianwei Lin (the 2018 ActivityNet Challenge winner of the Temporal Action Localization Task) for providing us with a sample submission file on the ActivityNet dataset.

What datasets can you use?

We provide augmented dataset files for the ActivityNet v1.3 (validation) and THUMOS14 dataset in the data folder. However, this tool is developed for any general dataset and any augmentation characteristics.

Who is behind it?

Contributor Contributor Contributor Advisor
Humam Alwassel Fabian Caba Victor Escorcia Bernard Ghanem

Do you want to contribute?

  1. Check the open issues or open a new issue to start a discussion around your new idea or the bug you found
  2. Fork the repository and make your changes
  3. Send a pull request