Skip to content
This repository has been archived by the owner on Oct 8, 2021. It is now read-only.

Latest commit

 

History

History
249 lines (194 loc) · 11.7 KB

README.md

File metadata and controls

249 lines (194 loc) · 11.7 KB
Raw

COSMOS: Catching Out-of-Context Misinformation using Self-Supervised Learning

GitHub license Read the Docs

COSMOS dataset consists of images and captions scraped from news articles and other websites designed for training and evaluation of out-of-context use of images. We refer readers to the paper for more details. To get access to the dataset, please fill out this form. We will provide you script to download the dataset. The official documentation for the project can be found here



Dataset Description

Dataset Statistics

COSMOS dataset consist of three splits : Training (160 K images), Validation (40 K images) and Test (1700 images). For training, we do not have/use out-of-context annotations. We only use these annotations in the end to evaluate our model. The dataset stats are listed below.

Table 1: Dataset stats.

Split # Images # Captions Context Annotation
Train 161,752 360,749 No
Valid 41,006 90,036 No
Test 1700 3400 Yes

Data Format

The COSMOS training, validation and test sets are provided as JSON (JavaScript Object Notation) text files with the following attributes for every data sample stored as a dictionary:

File Structure for train.json and val.json

{	"img_local_path": <img_path>, 
	"articles": [
                 { "caption": <caption1>, 
                   "article_url": <url1>, 
                   "caption_modified": <caption_mod1>,
                   "entity_list": <entity_list1>},
                   
                 { "caption": <caption2>,
                   "article_url": <url2>,
                   "caption_modified": <caption_mod2>,
                   "entity_list": <entity_list2>},

                 { "caption": <caption3>,
                   "article_url": <url3>,
                   "caption_modified": <caption_mod3>,
                   "entity_list": <entity_list3>},
                   
                  ......

				 ],
    "maskrcnn_bboxes": [ [x1,y1,x2,y2], [x1,y1,x2,y2], ... ]
}

Table 2: Attributes in Train/Validation files.

Key Description
img_local_path Source path in dataset directory for the image
articles List of dict containing metadata for every caption associated with the image
caption Original Caption scraped from the news website
article_url Link to the website image and caption scraped from
caption_modified Modified caption after applying Spacy NER (We used these caption as input to our model during experiments)
entity_list List that consists of mapping between modified named entities in the caption with the corresponding hypernym
maskrcnn_bboxes List of detected bounding boxes corresponding to the image. (x1,y1) refers to start vertex of the rectangle and (x2, y2) refers to end vertex of the rectangle

Note that for detecting bounding boxes, we used Detectron2 pretrained model linked here. We detect upto 10 bounding boxes per image.


File Structure for test.json

{	
        "img_local_path": <img_path>,
	"caption1": <caption1>,
	"caption1_modified": <caption1_modified>,
	"caption1_entities": <caption1_entities>,
	"caption2": <caption2>,
	"caption2_modified": <caption2_modified>,
	"caption2_entities": <caption2_entities>,
	"article_url": <article_url>,
	"label": "ooc/not-ooc",
	"maskrcnn_bboxes": [ [x1,y1,x2,y2], [x1,y1,x2,y2], ... ]
}

Table 3: Attributes in Test file.

Key Description
img_local_path Source path in dataset directory for the image
caption1 First caption associated with the image
caption1_modified Modified Caption1 after applying Spacy NER
caption1_entities List that consists of mapping between modified named entities in the caption1 with the corresponding hypernym
caption2 Second caption associated with the image
caption2_modified Modified Caption2 after applying Spacy NER
caption2_entities List that consists of mapping between modified named entities in the caption2 with the corresponding hypernym
article_url Link to the website image and caption scraped from
label Class label whether the two captions are out-of-context with respect to the image (1=Out-of-Context, 0=Not-Out-of-Context )
maskrcnn_bboxes List of detected bounding boxes corresponding to the image. (x1,y1) refers to start vertex of the rectangle and (x2, y2) refers to end vertex of the rectangle

Getting started

The code is well-documented and should be easy to follow.

  1. Source Code: $ git clone this repo and install the Python dependencies from requirements.txt. The source code is implemented in PyTorch so familarity with PyTorch is expected.

  2. Dataset: Download the dataset by filling out the form here.

  3. Visualize Dataset: It is difficult to view the dataset using only JSON file. Navigate to the directory dataset_visualizer and follow the instructions to visualize the dataset using a simple Python-Flask based web tool

  4. Train and Test For Image-Text Matching Task: This code is based on Detectron2 to extract features from objects present in the image. Please setup and install detectron2 first if you wish to use our feature detector for images. The minimal changes to be done to detectron2 source code to extract object features are added to detectron2_changes directory. Navigate to detectron2 source code directory and simply run patch -p1 ../detectron2 < 0001-detectron2-mod.patch. Consider setting up detectron2 inside this directory, it worked seamlessly for me without doing many changes.
    All the training parameters are configured via utils/config.py. Specify hyperparameters, text-embeddings, threshold values, etc in the config .py file. Model names are specifed in the trainer script itself. Configure these parameters according to your need and start training.
    To train the model, execute the following command: python trainer_scipt.py -m train
    Once training is finished, then to evaluate the model with Match vs No-Match Accuracy, execute the following command: python trainer_scipt.py -m eval

Evaluating

Test For Out-of-Context Detection Accuracy: Once training is over, then to evaluate the model for out-of-Context Detection task, specify model name in evaluate_ooc.py.

In order to reproduce our results execaute the following commands for each section.

  • Section 3.1: Differential Sensing
    export COSMOS_IOU=0.5
    export COSMOS_DISABLE_ISFAKE=1
    export COSMOS_DISABLE_RECT_OPTIM=1

    python evaluate_ooc.py
    
    unset $(env | sed -n 's/^\(COSMOS.*\)=.*/\1/p')
  • Section 3.2: Fake-or-Fact
    export COSMOS_IOU=0.5
    export COSMOS_DISABLE_ISOPPOSITE=1
    export COSMOS_DISABLE_RECT_OPTIM=1

    python evaluate_ooc.py
    
    unset $(env | sed -n 's/^\(COSMOS.*\)=.*/\1/p')
  • Section 3.3: Object-Caption Matching
    export COSMOS_IOU=0.5
    export COSMOS_WORD_DISABLE=1

    python evaluate_ooc.py
    
    unset $(env | sed -n 's/^\(COSMOS.*\)=.*/\1/p')
  • Section 3.4: IoU Threshold Adjustment
    export COSMOS_IOU=0.25
    export COSMOS_WORD_DISABLE=1
    export COSMOS_DISABLE_RECT_OPTIM=1

    python evaluate_ooc.py
    
    unset $(env | sed -n 's/^\(COSMOS.*\)=.*/\1/p')
  • Section 3.1 + 3.2 + 3.3 + 3.4: Proposed Method
    python evaluate_ooc.py

Environment Variables

These variables modify the behaviour of our evaluation method

Method specific variables

Variable Name Description
COSMOS_DISABLE_ISOPPOSITE Disables Section 3.1 Differential Sensing.
COSMOS_DISABLE_ISFAKE Disables Section 3.2 Fake-or-Fact.
COSMOS_DISABLE_RECT_OPTIM Disables Section 3.3 Object-Caption Matching.
COSMOS_IOU Setting it to "0.5" disabled Section 3.4, and "0.25" enables it.
COSMOS_WORD_DISABLE Disables Section 3.1 and 3.2 altogether.

Dataset specific variables

Variable Name Description
COSMOS_BASE_DIR Base directory.
COSMOS_DATA_DIR (optional) Data directory.
COSMOS_TARGET_DIR (optional) Target directory

Comparison specific variables

Variable Name Description
COSMOS_COMPARE Setting this to "1" will enable comparison with original paper.
COSMOS_COMPARE_LEVEL Choose between [0, 1, 2] (0 is default). Changes verbosity level.

Citation

If you find our dataset or paper useful for your research , please include the following citation:

@inproceedings{10.1145/3458305.3479968,
	author = {Akgul, Tankut and Civelek, Tugce Erkilic and Ugur, Deniz and Begen, Ali C.},
	title = {COSMOS on Steroids: A Cheap Detector for Cheapfakes},
	year = {2021},
	isbn = {9781450384346},
	publisher = {Association for Computing Machinery},
	address = {New York, NY, USA},
	url = {https://doi.org/10.1145/3458305.3479968},
	doi = {10.1145/3458305.3479968},
	numpages = {5},
	keywords = {RNN, fake, SBERT, BERT, differential sensing, IoU, Cheapfakes},
	location = {Istanbul, Turkey},
	series = {MMSys '21}
}