Short-video Face Parsing Challenge 2021

Powered by maadaa.ai open dataset

Overview
  • Description
  • Contestants
  • Task
  • Important Dates
  • Prizes

Short videos on fashion trends, hot spots, street interviews, public education, and creative advertising etc. have been extremely popular nowadays, while only a few people know about computer vision.

In this contest, we will do facial semantic segmentation. To detect facial attributes (eg. mouth, eyes and nose) at a pixel level, which is more challenging than face masking and keypoint.

In this competition, you’re challenged to train a face parsing model, which is able to detect every frame of the short video.

The competition is organized by maadaa.aiPIC (Person in Context),  and Beihang University, with technical support by Extreme Mart platform at CVPR 2021.

CVPR is the premier conference in the field of computer vision and pattern recognition organized by IEEE. Due to the continuing impact of the COVID, CVPR 2021 will be held online from June 19, 2021 to June 21, 2021.

The competition is open to everyone, including individuals, universities, research institutes and enterprises. One Contestant can only join one team and the maximum number of contestants for each team is 3.


The hosts are responsible for writing the test, staff who are involved in this procedure are prohibited from participation in the competition.

The co-organizers can participate in the contest, but result evaluation and prizes are not offered to these participants.

  • Contestants will receive the competition training data and validation data by email.

  • Contestants will then need to conduct algorithm training and development on their local computer or server.

  • Make the folder seg as a compressed file derectly. There should be no redundant directory in the compressed package【seg.zip】
    The directory structure should be as follow:
  • After the file is submitted successfully, the accuracy score and ranking of contestants on the list will release and the ranking will update once an hour.

  • March 30th, 10:00AM - Competition registration start time,  data set release
  • May 6th, 2021 10:00AM -Leaderboard Open
  • June 4th, 11:59PM - Registration deadline / Final  submission deadline
  • June 8th - Winners Announced
  • June 21st - Winners announced at CVPR workshop

All deadlines are based UTC+8 time zone on the corresponding day unless otherwise noted. The competition organizers reserve the right to update the contest timeline if they deem it necessary.

  • 1st Place: 10000RMB  + CVPR 2021 Workshop Announcement
  • 2nd Place: 5000RMB
  • 3rd Place: 3000RMB 
Schedule
  • 1. Registration
  • 2. Data Set Released
  • 3. Upload Test Results
  • 4. Result Evaluation
  • 5. Leaderboard Update
  • Registration Open: March 30th, 2021 10:00 AM(UTC+8)
  • Leaderboard Open: May 6th, 2021 10:00 AM(UTC+8)
  • Registration Deadline & Submission Deadline : June 4th, 2021 11:59 PM(UTC+8)
  • Data Set Release: March 30th, 2021 10:00 AM(UTC+8)
  • June 4th, 2021 11:59PM (UTC+8) - Final submission deadline.

    You can submit 3 times a day
  • Result Evaluation: June 8th, 2021
  • The leaderboard will be updated hourly.
Data
Dataset Description

The dataset consists of video images collected on the internet, in the format of PNG result images and JSON files. There is a total of 1500 videos, each video has 20 images (1 frame per second for each video).

 

  • Train set: 1200 videos
  • Validation set: 100 videos
  • Test set: 100 videos
 Samples
1.The Face Parse 

The labeled results are saved in the 8 bit grayscale PNG file with the same name as the frame image. The corresponding pixel value is the index value of the above label. For example, 0 represents the background pixel, 1 represents the facial skin pixel…, 16 means that the pixel point is glasses.

The face parse labels:

# Label
1
face_skin
2
left_eyebrow
3
right_eyebrow
4
left_eye
5
right_eye
6
nose
7
lower _lip
8
inner_mouth
9
upper_lip
10
hair
11
left_eye_shadow
12
right_eye_shadow
13
left_ear
14
right_ear
15
hat
16
glasses
17
else_skin
0
background
Licenses

maadaa.ai is providing this open dataset under the CC BY-NC 4.0 license.

Organizers

Yongsheng Zhang

Bin Wang

Zitian Wang

Si Liu

Sponsor
Other Competition