Person-Segmentation-Keras

Person segmentation with Keras (SegNet, Unet, etc.)

Dataset

Person segmentation

HumanParsing-Dataset is adopted in this repo.

Origin HumanParsing-Dataset contains 16+1 object classes. But in this repo, i just segment person which is a binary classification task.

I generate new label images by my self, you can download new label imgages at https://pan.baidu.com/s/1Y6bKUznsVc7xNWb9tqWaHA passwd: p8ks

Of course you can generate label images by yourself using convert_labels.py , gen_train_test_list.py.

I use 12706 images of HumanParsing-Dataset as training set, the remaining images as test set.

During training, i resize images with unchanged aspect ratio using padding, for details you can see this script.

Human parsing

Origin HumanParsing-Dataset contains 16+1 object classes.

background     0
hat            1
hair           2 
sunglass       3
upper-clothes  4
skirt          5
pants          6
dress          7
belt           8
left-shoe      9
right-shoe     10
face           11
left-leg       12
right-leg      13
left-arm       14
right-arm      15
bag            16
scarf          17

I have simplified the parsing task. Now it contains 4 + 1 classes.

background     0 (background, bag)
head           1 (hat, hair, sunglass, face, scarf)
upper body     2 (upper-clothes)
both hands     3 (left-arm, right-arm)
lower body     4 (skirt, pants, dress, belt, left-shoe, right-shoe, left-leg, right-leg)

New label images can be downloaded from this link https://pan.baidu.com/s/1jhqpOn8oBmiJiwzohhkfww passwd: gamc

Usage

All models are defined in 'models' directory.

An example for training Unet.

python train_segmentation.py --model='unet'

For visiualsizing the predictions, you can use the follow command

python predict.py --model='unet'

Results

Unet

Person segmentation

mIU: 0.8918

Origin images, ground truth images and predictions.

Val accuracy during training.

Val loss during training.

Human parsing

	Part	mIoU
Unet	head	0.66476
	upper body	0.48639
	both hands	0.27016
	lower body	0.66536
	mean	0.52167

Origin images, ground truth images and predictions.

SegNet

Todo

Reference

image-segmentation-keras

SegNet-Tutorial