not all pixels are equal difficulty-aware semantic...

13
Not All Pixels are Equal : Difficulty-aware Semantic Segmentation via Deep Layer Cascade Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, Xiaoou Tang Multimedia Lab, The Chinese University of Hong Kong

Upload: others

Post on 27-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Not All Pixels are Equal:Difficulty-aware Semantic Segmentation

via Deep Layer Cascade

Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, Xiaoou Tang

Multimedia Lab, The Chinese University of Hong Kong

Page 2: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

State-of-the-art Method (4 FPS)

Deep Layer Cascade (17 FPS)

Problem

Input Video

Page 3: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

State-of-the-art

State-of-the-art Method (4 FPS) Fully Convolutional Network

• Why Slow?

• Very Deep Backbone Network

• High Resolution Feature Map

Page 4: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Motivation

Image Easy Region Moderate Region Hard Region

Page 5: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Contemporary Model

Image Stem Reduction-A5×IRNet-A 10×IRNet-B Reduction-B 5×IRNet-C

pooling

Fully-Connected Softmax

Image Stem Reduction-A5×IRNet-A 10×IRNet-B Reduction-B 5×IRNet-C ConvI L3

Page 6: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Deep Layer Cascade

background

horse

person

car

unknown

Sta

ge 1

Image Stem Reduction-A5×IRNet-A

Conv

I

L1

Page 7: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Deep Layer Cascade

background

horse

person

car

unknown

Sta

ge 1

Sta

ge 2

Image Stem Reduction-A5×IRNet-A 10×IRNet-B Reduction-B

ConvConv

I

L1 L2

Page 8: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Deep Layer Cascade

background

horse

person

car

unknown

Sta

ge 1

Sta

ge 2

Sta

ge 3

Image Stem Reduction-A5×IRNet-A 10×IRNet-B Reduction-B 5×IRNet-C Conv

ConvConv

I

L1 L2

L3

Page 9: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Region Convolution

Convolution Region Convolution

M

Region Convolution with Residual

+

Page 10: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Performance

PASCAL VOC 2012

mIoU FPS (Backbone Network)

DPN 77.5 5.7

Adelaide 79.1 -

Deeplab-v2 79.7 7.1

LC(w/o COCO) 80.314.7

LC(with COCO) 82.7(PASCAL VOC 2012 Challenge test set)

Page 11: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Stage Visualization

(a) input image (e) ground truth(b) stage-1 (c) stage-2 (d) stage-3

background aeroplane person carbottle cat busunknown

Page 12: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

• Difficulty-Aware Learning Paradigm

input video stage-1

stage-2 stage-3

• Region Convolution Real-Time

• End-To-End Trainable Framework

Page 13: Not All Pixels are Equal Difficulty-aware Semantic ...luoping.me/project/layercascade/LayerCascade.pdf · Not All Pixels are Equal: Difficulty-aware Semantic Segmentation via Deep

Not All Pixels are Equal:Difficulty-aware Semantic Segmentation via Deep Layer Cascade

Thanks!

Code and models are available @

Project Page: http://personal.ie.cuhk.edu.hk/~lz013/projects/LayerCascade.html