hazeld / PseudoAdverbs

PseudoAdverbs

Datasets & Code for the CVPR 2022 paper 'How Do You Do It? Fine-Grained Action Understanding with Pseudo-Adverbs'

Datasets

The three datasets proposed in this work can be found in the datasets folder. Each dataset has its own folder with an annotations.csv file and an adverbs.csv file.

adverbs.csv contains the list of adverbs annotated in the dataset and their corresponding antonyms.

annotations.csv contains the YouTube video ID, the start and end time of the video clip and the action-adverb annotation. It has the following columns:

Column Name	Type	Example	Description
clip_id	string	video663	Unique id for the video clip corresponding to the original dataset
youtube_id	string	S7wF6S5ywo4	YouTube id for the full video
start_time	float	3.22	Value in seconds of the start time for the video clip
end_time	float	19.07	Value in seconds of the end time for the video clip
action	string	drive	The original action from the caption
adverb	string	rapidly	The original adverb from the caption
caption	string	a blue lamborghini drives rapidly down a busy street	The full caption for the video clip
clustered_adverb	string	drive	Annotated adverb
clustered_action	string	quickly	Annotated action

Videos can be downloaded with python utils/download_videos.py datasets/<dataset_name>/annotations.csv <download_dir>

Extracted features can be downloaded here: https://drive.google.com/drive/folders/1gdF_Hp7hBmlRHqs9TG0pehZzSzaZw5Bc?usp=sharing

Code

Trained models can be downloaded here: https://drive.google.com/file/d/1BqrzjdFQ5bqrhlWakiNQdi_5NE7h64Nb/view?usp=sharing

For models trained on VATEX we provide models trained for action embedding for the first 200 epochs here: https://drive.google.com/file/d/13aTARYglJfHa11SVRO9BSC6V3eEyIBdS/view?usp=sharing

Below we give the commands needed to train and test models for the different tasks and datasets.

Seen Compositions

VATEX Adverbs

Train

python train.py --data-dir splits/with_unlabelled/seen_compositions/vatex_adverbs_5/ --train-feature-dir data/VATEX_Adverbs/features/ --test-feature-dir data/VATEX_Adverbs/features/ --checkpoint-dir checkpoints/vatex_adverbs_5_pseudo_adverbs/ --modality both --unlabelled-ratio 19 --load checkpoints/action_pretraining/vatex_adverbs_5_action_pretraining --pseudo-label-threshold 0.6 --smoothing 0.1 --adaptive-threshold

Test

python test.py --unlabelled-ratio 1 --data-dir splits/with_unlabelled/seen_compositions/vatex_adverbs_5/ --test-feature-dir data/VATEX_Adverbs/features/ --load checkpoints/vatex_adverbs_5_pseudo_adverbs/ckpt_E_1000

HowTo100M Adverbs

Train

python train.py --data-dir splits/with_unlabelled/seen_compositions/howto100m_adverbs_10/ --train-feature-dir data/HowTo100M_Adverbs/features/ --test-feature-dir data/HowTo100M_Adverbs/features/ --checkpoint-dir checkpoints/howto100m_adverbs_10_pseudo_adverbs/ --modality both --unlabelled-ratio 9 --pseduo-label-threshold 0.6 --smoothing 0.1 --adaptive-threshold --t_train 20 --t_test 20 --num-pseudo-labelled 3 --pseudo-action-pretraining --max-epochs 1000

Test

python test.py --unlabelled-ratio 1 --t_test 20 --data-dir splits/with_unlabelled/seen_compositions/howto100m_adverbs_10/ --test-feature-dir data/HowTo100M_Adverbs/features/ --load checkpoints/howto100m_adverbs_10_pseudo_adverbs/ckpt_E_1000 --instance-av

Unseen Compositions

Train

python train.py --data-dir splits/with_unlabelled/unseen_compositions/vatex/ --train-feature-dir data/VATEX_Adverbs/features/ --test-feature-dir data/VATEX_Adverbs/features/ --checkpoint-dir checkpoints/vatex_adverbs_unseen_pseudo_adverbs/ --modality both --unlabelled-ratio 1 --load checkpoints/action_pretraining/vatex_adverbs_unseen_action_pretraining --pseudo-label-threshold 0.6 --smoothing 0.1 --adaptive-threshold

Test

python test.py --unlabelled-ratio 1 --data-dir splits/with_unlabelled/unseen_compositions/vatex/ --test-feature-dir data/VATEX_Adverbs/features/ --load checkpoints/vatex_adverbs_unseen_pseudo_adverbs/ckpt_E_1000

Unseen Domains

Train

python train.py --data-dir splits/with_unlabelled/unseen_domains/vatex2msrvtt/ --train-feature-dir data/VATEX_Adverbs/features/ --unlabelled-feature-dir data/MSR-VTT_Adverbs/features/ --test-feature-dir data/MSR-VTT_Adverbs/features/ --checkpoint-dir checkpoints/vatex2msrvtt_pseudo_adverbs/ --modality both --unlabelled-ratio 1 --load checkpoints/action_pretraining/vatex2msrvtt_pseudo_adverbs_action_pretraining --pseudo-label-threshold 0.6 --smoothing 0.1 --adaptive-threshold

Test

python test.py --unlabelled-ratio 1 --data-dir splits/with_unlabelled/unseen_domains/vatex2msrvtt/ --test-feature-dir data/MSR-VTT_Adverbs/features/ --load checkpoints/vatex2msrvtt_pseudo_adverbs/ckpt_E_1000

Citation

If you find this work helpful in your research, please cite:

@inproceedings{doughty2022how,
    author    = {Doughty, Hazel and Snoek, Cees G. M.},
    title     = {{H}ow {D}o {Y}ou {D}o {I}t? {F}ine-{G}rained {A}ction {U}nderstanding with {P}seudo-{A}dverbs},
    booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2022}
}