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:
@inproceedings{alwassel_2018_detad,
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)},
month={September}
year={2018}
}
Install conda, create the environment from the environment.yml
file, and then activate the environment:
conda env create -f environment.yml
source activate detad
Run the false positive analysis (the other two analyses are run in a similar way):
cd src
python false_positive_analysis.py --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/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 ActivityNet v1.3 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.
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.
Contributor | Contributor | Contributor | Advisor |
Humam Alwassel | Fabian Caba | Victor Escorcia | Bernard Ghanem |
- Check the open issues or open a new issue to start a discussion around your new idea or the bug you found
- Fork the repository and make your changes
- Send a pull request