aia39 / activity-classification-from-first-person-office-videos-with-visual-privacy-protection

This is the official repository of the paper titled 'Activity Classification from First Person Office Videos with Visual Privacy Protection' which is accepted in IC4IR 2021.

Activity-Classification-from-First-Person-Office-Videos-with-Visual-Privacy-Protection

This is the official repository of the paper titled 'Activity Classification from First Person Office Videos with Visual Privacy Protection' accepted in IC4IR 2021. Paper link. This is an open source activity (18 classes) recognition from human concerned privacy protected video codebase. We provide both the training and testing scripts along with model weight for any comparison purposes. This repository includes implementations of the following methods

• Privacy protected video generation using Mask R-CNN (backbone Inception ResNet V2)
• Video classification on raw rgb frames.
• Video classification on blurred (yielded from privacy protection scheme) rgb frames.
• Video classification from mixed training (leveraging both the rgb and blurred video learning)

Introduction

The goal of our approach is to provide a high-performing, visually privacy aware, modular codebase which provides user to make the most use of any module of our work. In the visual privacy protection scheme Inception ResNet V2 is used as backbone and Mask-RCNN is used to detect specific masks. Firstly, inspecting the whole FPV-O dataset which was provided in Video and Image Processing Cup 2019 we identified privacy violating sensitive objects for each class. All the sensitive objects are listed in below table.

Distribution of Sensitive Objects in Different Classes

Activity Class	Sensitive Objects
Chat	Person, digital screen, keyboard, laptop, book
Clean	Person
Drink	Person, Digital Screen
Dryer	Person, Toilet
Machine	Person, Digital Screen
Microwave	Person
Mobile	Mobile, Digital Screen, Keyboard, Book, Person
Paper	Book, Laptop, Digital Screen, Keyboard, Person
Print	Person, Digital Screen, Laptop, Book
Read	Digital Screen, Keyboard, Person, Laptop
Shake	Digital Screen, Laptop, Person
Staple	Book, Digital Screen, Keyboard, Person
Take	Digital Screen, Person
Typeset	Digital Screen, Keyboard, Person, Laptop
Walk	Digital Screen, Keyboard, Laptop
Wash	Toilet, Person
Whiteboard	Person, Book
Write	Digital Screen, Keyboard, Person, Laptop, Book

In the activity recognition section, we developed an ensembled based LSTM model with attention which trained on mixed dataset (described in the paper). Ensemble on the models are:

• ResNeXt 101 + Attention
• WideResNet 101 + Attention
• DenseNet + Attention
• WideResNet 101

Folder Directory Hierarchy (Visual Privacy Protection module)

test_directory_folder/
        video1.mp4
        video2.mp4
        ...
        

Output will be saved in protected_directory_folder

Folder Directory Hierarchy (Activity Recognition module)

src/
    Activity_Recognition/
        data/
        log/
        test_directory_folder/
        model_checkpoint/

        ...

Installation (Modular Implementation)

For Visual Privacy Protection module,

1. Install dependencies

   conda install -c anaconda tensorflow-gpu
   conda install -c anaconda pillow
   conda install -c anaconda opencv
   conda install -c anaconda matplotlib

2. Download pre-trained weights,classes names and related file from the Tensorflow model zoo. Download 'mask_rcnn_inception_resnet_v2_atrous_coco' from 'COCO-trained models' table. Put the .rar file in ['object_detection'](https://github.com/aia39/Activity-Classification-from-First-Person-Office-Videos-with-Visual-Privacy-Protection/src/Privacy protected video generation/object_detection) folder. You can also download other models which gives mask as output. We select inception_resnet_V2 as it gives better result though computationally expensive.

3. Create two folders in 'object_detection' folder 'test_directory_folder', 'protected_directory_folder' respectively for test videos and protected videos.

4. Run the [py](https://github.com/aia39/Activity-Classification-from-First-Person-Office-Videos-with-Visual-Privacy-Protection/src/Privacy protected video generation/object_detection/protected.py) file in command window from 'object_detection' folder to generate masked video/frame.

  python protected.py

For Activity Recognition module,

1. Install dependencies

   conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
   conda install -c conda-forge tqdm
   conda install -c conda-forge av
   conda install -c anaconda pillow
   conda install numpy matplotlib scikit-learn

2. To test videos, place the videos in the test_directory_folder and run this code. Before using the script place the weight file in the model_checkpoint folder. (Weights will be provided soon)

3. To train your own model, we have provided here with four models and their corresponding four training script named as train{i}.py where i represents the different models.

Demo

See the youtube video for the quality of visually privacy protected video.

Acknowledgement

Thanks to mhealth lab for providing world class research facility.

Activity-Classification-from-First-Person-Office-Videos-with-Visual-Privacy-Protection

If you find 'Activity-Classification-from-First-Person-Office-Videos-with-Visual-Privacy-Protection' useful in your research, please use the following BibTeX entry for citation.

@inproceedings{ghosh2022activity,
  title={Activity Classification from First-Person Office Videos with Visual Privacy Protection},
  author={Ghosh, Partho and Istiak, Md and Rashid, Nayeeb and Akash, Ahsan Habib and Abrar, Ridwan and Dastider, Ankan Ghosh and Sushmit, Asif Shahriyar and Hasan, Taufiq and others},
  booktitle={Proceedings of International Conference on Fourth Industrial Revolution and Beyond 2021},
  pages={371--384},
  year={2022},
  organization={Springer}
}