CVPR2021 High-resolution Human Parsing Challenge

High-resolution Human Parsing Challenge

Overview

This challenge aims to recognize human parts (19 semantics in total) within high-resolution images by learning with low-resolution ones. To date, there is little existing data for such a challenge thus, we annotated 10,500 single-person images (training/validation/testing: 6,000/500/4,000) with an average resolution of 3950 by 2200. In addition to the provided high-resolution images, we also welcome off-the-shelf low-resolution datasets such as LIP and Pascal-Person-Part for pre-training. This new track poses a new task of learning from imperfect data, transferring the learned knowledge from low-resolution images to high-resolution images.

maadaa.ai is collaborating with Learning from Limited and Imperfect Data (L2ID) to host this challenge at CVPR 2021.

https://l2id.github.io/challenge_localization.html

April 1, 2021 – Start Date
May 15, 2021 – End Date

Data

Dataset Description

Image Quantity：5000 pcs
Image Size: 2000*3000 pixels
Geographies Covered: Asia, East Asia, Europe, America, Africa
Gender ratio: 1:1
Age: Teens to middle-aged adults
Hairstyle: Long hair, short hair, curly hair, straight hair, shoulder-length hair, and others.
Posture: Standing, walking, running, sitting accounted for 70%, unique postures such as squatting and cross-legged accounted for 30%.
Scenes: Street, playgrounds, shopping malls, office, environment under security cameras etc.
Image format: JPG
Data output format: Mask, PNG, JSON
Data accuracy: More than 98%.

Samples

Images:

Annotation samples:

The labels:
Anything not included in these 19 labels is background.

#	Label
1	hat
2	hair
3	sunglasses
4	shirt
5	dress
6	skirt
7	coat
8	socks
9	pants
10	glove
11	scarf
12	face
13	skin
14	left arm
15	right arm
16	left leg
17	right leg
18	left shoe
19	right shoe
20	background