Mask Detection Challenge

Detect Masks in the Wild!

TBA Prize Money



After initial debate, the utility of masks during the Covid-19 pandemic seems widely accepted now. The dominant scientific opinion says masks are very useful, and even relatively simple home-made masks can offer a great degree of protection against the novel coronavirus. With several countries like Singapore, Japan, states, and cities in the United States like Texas and New York making wearing masks mandatory for the public there comes a need for an automated system that is capable of helping authorities maintain guidelines which is where we can hope to turn to computer vision for a solution.

πŸ‘ The Vision

This challenge was created with the vision of developing a real-time mask detection system available for public use, to help public health officials all over the world. We hope that the models developed here by the AI/ML community enable developers around the globe to be able to use and deploy the same to build systems that would be capable to withstand the demands of a real-time, real-world use case. In particular, it would help factories ensure mask compliance is followed, help ensure safety for visitors in containment zones or hospitals where it is crucial for such measures to be taken, and so on. The applications are boundless and are of urgent need in this critical time.

Apart from these applications, this challenge hopes to serve as an educational tool that provides beginners with comprehensive starter kits and appropriate resources to help jumpstart their learning on object detection. The starter kit will be released soon and we encourage the community to release similar resources and help make this a great learning experience for all.

Note: If you wish to contribute to the baseline or add your own please follow these instructions or feel free to contact the challenge organizers.   

πŸ’Ό Problem Statement

The goal of this challenge is to train object detection models capable of identifying the location of masked faces in an image as well as the location of unmasked faces in the image. These detectors should be robust to noise and provide as little room as possible to accommodate false positives for masks due to the potentially dire consequences that they would lead to. Ideally, they should be fast enough to work well for real-world applications, something we hope to focus on in future rounds of the competition.

Code coming soon! πŸ˜„

πŸ’Ύ Dataset

The dataset that would be used is a growing dataset from a combination of many sources, either hand-labeled or pseudo-labeled. The dataset would contain annotations for masked faces, unmasked faces as well as some images that don't have any annotations as would be the case in the real-world scenario.

Note: If you wish to contribute to the dataset please follow these instructions or feel free to contact the challenge organizers.

This dataset would be released soon.


Check out this repository that shows a live web-cam-demo of a sample model in action!

πŸ“ Files

To be released soon

There are two annotation files:

- train.json

- test.json

Both of the files contain annotations for images in the train/ and the test/ folders and follow the MS COCO annotation format:


"id": int, "image_id": int, "category_id": int, "segmentation": RLE or [polygon], "area": float, "bbox": [x,y,width,height], "iscrowd": 0 or 1,




"id": int, "name": str, "supercategory": str,


πŸš€ Submission

 Submission Instructions :

You will be required to submit your predictions as a json file that is in accordance to the MS COCO.

For detection with bounding boxes, please use the following format:


"image_id": int, "category_id": int, "bbox": [x,y,width,height], "score": float,




{"image_id": int, "bbox": [ float, float, float, float], "score": float, "category_id": int },

{"image_id": int, "bbox": [ float, float, float, float], "score": float, "category_id": int },




πŸ–Š Evaluation Criteria

IoU (Intersection Over Union)

IoU measures the overall overlap between the true region and the proposed region. Then we consider it a True detection, when there is atleast half an overlap, or when IoU > 0.5.


Then we can define the following parameters :

Precision (IoU > 0.5) :   

Recall (IoU > 0.5) :   

The final scoring parameters AP_{IoU > 0.5} and AR_{IoU > 0.5} are computed by averaging over all the precision and recall values for all known annotations in the ground truth.

A further discussion about the evaluation metric can be found here.

The evaluation code can be found here.

πŸ“± Contact

- Shraddhaa Mohan

- Rohit Midha

- Sharada Mohanty