Skip to content

hellopipu/person_seg

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Person Segmentation

Introduction

we implement two versions of the model, one of which is a real-time high-precision version on GPU platform and the other is a real-time version on CPU platform.

model   input size Val iou (Supervisely)    Val iou (Whole)   speed (FPS) Size (MB)
Linknet_res18     640 88% 78% 100 (P40)  44
Unet++     256 56% --- 24 (CPU)   1.58

Datasets

 Dataset Train Val
Supervisely     4757 526
VOC_person    401 44
COCO_person     58023 6411
COCO_negtive     54172 0
VIP     16702 1846
TRIMODAL     5722 0
Total 139777 8827

Models

1). Linknet

we use linknet_res18 as the model to implement the real-time high-precision version on GPU platform

linknet

2). Unet++

we use unet++ as the model to implement the real-time version on CPU platform. Notice that the model is not the same as the origin unet++ mentioned in paper. while testing , we only keep X^00 , X^01 , X^10 to do the inference .

unet++

Demo

download the pretrained linknet model[BaiduNet](code: ftfe), and run the command below to try the demo

python ./demo/demo.py 

linknet

 

Releases

No releases published

Packages

No packages published

Languages