Back to all updates

almost 5 years ago

Problem Statement Released

Face Mask Detection in a Crowd

Problem Statement:

Given a video sequence, you are required to detect all the faces in every frame, classify each face as masked or non-masked, uniquely identify each person and track the duration for which each person is masked and non-masked.

Tasks:

  1. Create/pick multiple video clips capturing moving people with some of them putting masks on and some without them. The video should satisfy the following constraints: a. Duration should be of 120 seconds b. Resolution should be of at least 640 X 480 (@ 15 fps or more) with RGB888 format c. The video clip should contain at least 5 unique faces d. Video should consist minimum of 3 unique faces in the frame, over a minimum duration of 5 seconds with in the captured 2 Min video to qualify.
  2. Develop a solution that can detect several distinct faces in a frame and ID/tag them
  3. Track each unique face distinctly throughout the duration of the video and capture its mask status for every frame it has been a part of.
  4. Generate a corresponding output video capturing bounding boxes for each unique face in every frame as specified below. a. Face is masked – Green bounding box b. Face is not masked – Red bounding box c. Face Id(i.e unique person id) is embedded in the bounding box
  5. Summarize the time duration statistics for each unique face and generate the output in a CSV file as per the format detailed in Submission Guidelines.

Sample Video Frame:

enter image description here

Submission Guidelines:
  1. Submission should contain a minimum of 5, 2-minute input video sequences with a minimum resolution of 640x480 (@ 15 fps or above) and the same is applicable for output videos as well.
  2. Fully functional Python Scripts used for Training and Testing (module should contain setup.py taking care of all pip dependencies) Ref: https://godatadriven.com/blog/a-practical-guide-to-using-setup-py/
  3. Trained model(s) with input data-sources.
  4. Output video sequence highlighting masked faced in green and non-masked faces in red bounding boxes. Each bounding box should also have a unique ID/label displayed for each face in every frame.
  5. Flow chart of the solution containing training strategies, pre and post processing methods, tracking algorithms.
  6. Document list frameworks, libraries being used and instructions to run the inference
  7. A text file (for each output video) capturing the following details for each unique face id/person, a. Mask On/Off Stats: List of Entry & Exit Timestamps in the respective videos Note: Timestamp for the first frame in the video is 0.000 seconds
  8. CSV filed in format specified below with the required details (5 videos)

S2

Evaluation:

  • Each solution will be tested against their own 5 input videos submitted along with the solution.
  • Testing against hidden dataset for accuracy.
  • Inference latency will also be a key component of the evaluation criteria.
  • Plagiarism should be avoided

FAQs:

  1. Can we use any dataset for training?

    a. Yes, you are free to use any dataset of your choice

  2. Is the number of training samples fixed?

    a. No. Any number of training samples can be used

  3. Which framework to use?

    a. Any Python friendly framework is accepted.

  4. What is the number of submissions allowed for each team?

    a. Only one submission per team is allowed

  5. What are the start and end times of the hackathon?

    a. Teams sizes can vary from 1 to 4

  6. What are the limits on team size?

    a. Megathon 2021 starts on 22nd October 5:30PM IST and ends 48 hours later on 24th October 5:30PM IST

  7. Which video formats are accepted for submission?

    a. MP4