Self-Supervised Multi-Task Procedure Learning from Instructional Videos

Overview

This repository contains the implementation of Self-Supervised Multi-Task Procedure Learning from Instructional Videos.

---

Prerequisites

• Python 3.x
• PyTorch 1.x.x
• sklearn
• matplotlib
• skimage
• scipy

---

Data Preparation

CrossTask

1. Please download and extract the annotations and video links of CrossTask to ./data/CrossTask via https://www.di.ens.fr/~dzhukov/crosstask/crosstask_release.zip

2. To crawl videos in CrossTask from YouTube, please run:

python ./download_youtube/download_youtube_CrossTask.py

3. To extract and subsample video frame feature, please run:

python ./extract_feature/extract_img_frame_CrossTask.py
python ./extract_feature/extract_feature_cats_CrossTask_subsample.py

• We use the original training/testing split from CrossTask dataset

ProceL

1. Please download and extract the annotations and video links of ProceL to ./data/ProceL via https://drive.google.com/file/d/1b8PoZlYeNMP3PieJ3_80KkeibaCmmljS/view?usp=sharing

2. To crawl videos in ProceL from YouTube, please run:

python ./download_youtube/download_youtube_ProceL.py

3. Please run the following preprocessing script to gather annotation file in ProceL:

python ./preprocess/rename_gather_mat_file_ProceL.py

4. To extract and subsample video frame feature, please run:

python ./extract_feature/extract_img_frame_ProceL.py
python ./extract_feature/extract_feature_cats_ProceL.py
python ./extract_feature/extract_feature_cats_ProceL_subsample.py

• Information about the training split of ProceL is in ./data_partition/ProceL_train_partition.csv

---

Training

[CrossTask] Task-Specific Setting

1. To train the task-specific model on CrossTask dataset, please run:

chmod +x ./chain_experiment_script/CrossTask/CrossTask_same_cat_ss.sh
./chain_experiment_script/CrossTask/CrossTask_same_cat_ss.sh

[CrossTask] Multi-Task Setting

1. To train the multi-task model on CrossTask dataset, please run:

chmod +x ./chain_experiment_script/CrossTask/CrossTask_all_cat_experiments.sh
./chain_experiment_script/CrossTask/CrossTask_all_cat_experiments.sh

[ProceL] Multi-Task Setting

1. To train the multi-task model on ProceL dataset, please run:

chmod +x ./chain_experiment_script/ProceL/all_cat_experiments.sh
./chain_experiment_script/ProceL/all_cat_experiments.sh

Pretrained Models

For ease of reproducing the results, we provided the pretrained models for:

Dataset	Setting	Model
CrossTask	K=15 (Teacher model)	download
CrossTask	K=15 (Student model)	download
Procel	K=15 (Teacher model)	download
Procel	K=15 (Student model)	download

---

Citation

If this code is helpful for your research, we would appreciate if you cite the work:

@article{Elhamifar-MultiTaskProcedureLearning:ECCV20,
  author = {E.~Elhamifar and D.~Huynh},
  title = {Self-Supervised Multi-Task Procedure Learning from Instructional Videos},
  journal = {European Conference on Computer Vision},
  year = {2020}}