Accurate Estimation of Body Height from a Single Depth Image   via a Four-Stage Developing Network

Dataset Introduction

  We use a Kinect camera to create a human body dataset with 2136 RGB-D images which consists of 10 postures, including upright, walking, sitting, bending, arms raising slightly, unrolling arms, arms over head, waving hands, clapping and having a waistline. There are 14 volunteers in our dataset. They can stand anywhere with arbitrary clothes. Their heights ranges from 158cm to 184cm, which covers a wide range of height. The following figure shows some examples:

Composition and Division of Our Dataset

  Our entire dataset is divided into three parts: part A, B and C. Part A is the training set with 1707 images. partB and C together forms the test set with totally 429 images.  We first divide 1767 images from 12 people into two parts: part A and B . Part A has 1707 images to form the training set. PartB, denoted as Familiar-test, has 60 images and is part of the test set. Network may recognize the identity information from the image of Part B.
 The 369 images of all the other volunteers are part C, denoted as Strange-test. It is impossible for network to recognize any identity information from part C since none of them appears in the training set. It will be very easy to determine whether the network has learned identity information or body height under such a data set configuration. It only need to compare the difference of accuracy between part B and C.

Intermediate Representation Generation Method

 Torso Information T: We first train an FCN network using the PASCAL VOC dataset to recognize human from human body from our color images, then we do manual annotation to retain only the torso information T.
 Label Image L: Next the joint points are automatically recognized and then we do some manual corrections to get the four part segmentation image as Label Image L.

Citation

If you use the dataset or code, please cite the following work:

@InProceedings{Yin_2020_CVPR,
author = {Yin, Fukun and Zhou, Shizhe},
title = {Accurate Estimation of Body Height From a Single Depth Image via a Four-Stage Developing Network},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

Download Link

Please contact us by email.