This repository is the official implementation of Temporal and cross-modal attention for audio-visual zero-shot learning.
Install all required dependencies into a new virtual environment via conda.
conda env create -f TCAF.yml
We base our datasets on the AVCA repository. The dataset structure is identical to AVCA and the dataset folder is called avgzsl_benchmark_non_averaged_datasets
. The only difference is that we use temporal features instead of averaged features. We provide our temporal C3D/VGGish features to download below.
In order to extract the C3D/VGGish features on your own, run the scripts in the /cls_feature_extraction
as follows:
python3 cls_feature_extraction/get_features_activitynet.py
python3 cls_feature_extraction/get_features_ucf.py
python3 cls_feature_extraction/get_features_vggsound.py
Given the files extracted by the above scripts, run the following command to obtain the cls features:
python3 splitting_scripts_cls/create_pkl_files_cls.py --dataset_name DATASET_NAME --path_original_dataset PATH_ORIGINAL_DATASET --path_splitted_dataset PATH_SPLITTED_DATASET
arguments:
--dataset_name: Name of the dataset
--path_original_dataset: the path of the dataset where the above scripts (those in ```cls_feature_extraction```) have extracted the dataset
--path_splitted_dataset: the path where to put the dataset after it is processed in the right way.
Moreover, we adapted the SeLaVi implementation from the AVCA repository in order to extract temporal features and to make extraction more parallelizable. For obtaining the SeLaVi features we used the following commands:
python3 selavi_feature_extraction/get_clusters.py \
--root_dir <path_to_raw_videos> \
--weights_path <path_to_pretrained_selavi_vgg_sound.pth> \
--mode train \
--pretrained False \
--aud_sample_rate 44100 \
--use_mlp False \
--dataset {activity,ucf,vggsound} \
--headcount 2 \
--exp_desc <experiment_description> \
--output_dir <path_to_save_extracted_features> \
--batch_size 1 \
--workers 0
python3 selavi_feature_extraction/merge_features_selavi.py
python3 splitting_scripts_main/create_pkl_files_selavi.py --dataset_name DATASET_NAME --path_original_dataset PATH_ORIGINAL_DATASET --path_splitted_dataset PATH_SPLITTED_DATASET
arguments:
--dataset_name: Name of the dataset
--path_original_dataset: the path of the dataset where the above scripts (those in ```selavi_feature_extraction```) have extracted the dataset
--path_splitted_dataset: the path where to put the dataset after it is processed in the right way.
You can download our temporal supervised (C3D/VGGish) features of all three datasets here:
We additionally provide temporal self-supervised (SeLaVi) features, which have been pretrained in self-supervised manner on VGGSound:
Since the VGGSound dataset is also used for the zero-shot learning task, we recommend the usage of supervised (C3D/VGGish) features instead of SeLaVi.
The features should be placed inside the avgzsl_benchmark_non_averaged_datasets
folder:
unzip [DATASET].zip -d avgzsl_benchmark_non_averaged_datasets/
In order to train the model run the following command:
python3 main.py --cfg CFG_FILE --root_dir ROOT_DIR --log_dir LOG_DIR --dataset_name DATASET_NAME --run all
arguments:
--cfg CFG_FILE is the file containing all the hyperparameters for the experiments. These can be found in ```config/best/X/best_Y.yaml``` where X indicate whether you want to use cls features or main features. Y indicate the dataset that you want to use.
--root_dir ROOT_DIR indicates the location where the dataset is stored.
--dataset_name {VGGSound, UCF, ActivityNet} indicate the name of the dataset.
--log_dir LOG_DIR indicates where to save the experiments.
--run {'all', 'stage-1', 'stage-2'}. 'all' indicates to run both training stages + evaluation, whereas 'stage-1', 'stage-2' indicates to run only those particular training stages
Evaluation can be done in two ways. Either you train with --run all
which means that after training the evaluation will be done automatically, or you can do it manually.
For manual evaluation run the following command:
python3 get_evaluation.py --cfg CFG_FILE --load_path_stage_A PATH_STAGE_A --load_path_stage_B PATH_STAGE_B --dataset_name DATASET_NAME --root_dir ROOT_DIR
arguments:
--cfg CFG_FILE is the file containing all the hyperparameters for the experiments. These can be found in ```config/best/X/best_Y.yaml``` where X indicate whether you want to use cls features or main features. Y indicate the dataset that you want to use.
--load_path_stage_A will indicate to the path that contains the network for stage 1
--load_path_stage_B will indicate to the path that contains the network for stage 2
--dataset_name {VGGSound, UCF, ActivityNet} will indicate the name of the dataset
--root_dir points to the location where the dataset is stored
The trained models can be downloaded from here.
Method | VGGSound-GZSL | UCF-GZSL | ActivityNet-GZSL |
---|---|---|---|
Attention fusion | 4.95 | 24.97 | 5.18 |
Perceiver | 4.93 | 34.11 | 6.92 |
CJME | 3.68 | 28.65 | 7.32 |
AVGZSLNET | 5.26 | 36.51 | 8.30 |
AVCA | 8.31 | 41.34 | 9.92 |
TCAF | 8.77 | 50.78 | 12.20 |
Method | VGGSound-GZSL | UCF-GZSL | ActivityNet-GZSL |
---|---|---|---|
Attention fusion | 3.37 | 20.21 | 4.88 |
Perceiver | 3.44 | 28.12 | 4.47 |
CJME | 3.72 | 29.01 | 6.29 |
AVGZSLNET | 4.81 | 31.51 | 6.39 |
AVCA | 6.91 | 37.72 | 7.58 |
TCAF | 7.41 | 44.64 | 7.96 |
src
- Contains the code used throughout the project for dataloaders/models/training/testing.
c3d
- Folder contains the code for the C3D network.
audioset_vggish_tensorflow_to_pytorch
- Contains the code which is used to obtain the audio features using VGGish.
cls_feature_extraction
- Contains the code used to extract the C3D/VGGish features from all 3 datasets.
selavi_feature_extraction
- Contains the code used to extract the SeLaVi features.
splitting_scripts_{cls,main}
- Contains files from spltting our dataset into the required structure.
If you find this code useful, please consider citing:
@inproceedings{mercea2022tcaf,
author = {Mercea, Otniel-Bogdan and Hummel, Thomas and Koepke, A. Sophia and Akata, Zeynep},
title = {Temporal and cross-modal attention for audio-visual zero-shot learning},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2022}
}
@inproceedings{mercea2022avca,
author = {Mercea, Otniel-Bogdan and Riesch, Lukas and Koepke, A. Sophia and Akata, Zeynep},
title = {Audio-visual Generalised Zero-shot Learning with Cross-modal Attention and Language},
booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2022}
}