Code for the paper A Flexible Model for Training Action Localization with Varying Levels of Supervision, NIPS 2018
Created by Jean-Baptiste Alayrac and Guilhem Chéron at INRIA, Paris.
The webpage for this project is available here. It contains link to the paper, and other material about the work. This code reproduces the results presented in Table 1 of the paper for our method.
Our code is released under the MIT License (refer to the LICENSE file for details).
If you find this code useful in your research, please, consider citing our paper:
@InProceedings{actoraction18, author = {Ch'eron, Guilhem and Alayrac, Jean-Baptiste and Laptev, Ivan and Schmid, Cordelia}, title = {A Flexible Model for Training Action Localization with Varying Levels of Supervision}, booktitle = {Neural Information Processing Systems (NIPS)}, year = {2018} }
We run the code under python 2.7 with the following dependencies:
- numpy
- tqdm
- mosek (for which you'll need a license)
- pickle
- scikit-learn
- scipy
- Clone this repo and go to the generated folder
git clone https://github.com/jalayrac/weak_action_loc.git
cd weak_action_loc
- Download and unpack the preprocessed features needed for the desired experiment in the data folder:
-
UCF101-24 (11 GiB)
mkdir -p data cd data wget https://www.di.ens.fr/willow/research/weakactionloc/UCF101-24.tar.gz tar -xzvf UCF101-24.tar.gz cd ..
-
DALY (48 GiB)
mkdir -p data
cd data
wget https://www.di.ens.fr/willow/research/weakactionloc/DALY.tar.gz
tar -xzvf DALY.tar.gz
cd ..
- To obtain the results, you need to first run the traning code to obtain the parameters of the model, and then run the evaluation code:
Dataset | Video level | Shot level | Temporal point | Temporal | Temporal + spatial points | 1 BB | Temp. + 1 BB | Temp. + 3 BBs | Fully supervised |
---|---|---|---|---|---|---|---|---|---|
UCF101-24 | script | - | script | script | script | script | script | script | script |
DALY | script | script | - | script | - | script | script | script | - |
NB: we provide the calibration files (threshold values) that were obtained by validation as described in the paper. We will release the code for this calibration in a future release.