MSCOCO Keypoints Challenge 2018 Megvii (Face++) Team members: - - PowerPoint PPT Presentation

▶
mscoco keypoints
SMART_READER_LITE
LIVE PREVIEW

MSCOCO Keypoints Challenge 2018 Megvii (Face++) Team members: - - PowerPoint PPT Presentation

Jul 25, 2023 252 likes •581 views

MSCOCO Keypoints Challenge 2018 Megvii (Face++) Team members: Qixiang Peng Zhicheng Wang Wenbo Li Binyi Yin Tianzi Xiao Yuming Du Zeming Li Jian Sun Gang Yu Xiangyu Zhang Megvii (Face++) Results on COCO 2018 Year 2016 2017 2018

slide-1
SLIDE 1

MSCOCO Keypoints Challenge 2018

Megvii (Face++)

slide-2
SLIDE 2

Megvii (Face++)

Team members:

Zhicheng Wang Gang Yu Zeming Li Jian Sun Xiangyu Zhang Wenbo Li Tianzi Xiao Binyi Yin Qixiang Peng Yuming Du

slide-3
SLIDE 3

Results on COCO 2018

Year 2016 2017 2018 mmAP 60.5 72.1 76.4(ours)

Results on COCO test challenge recent years

slide-4
SLIDE 4

Results on COCO 2018

test_challenge test_dev mini_validation Ensemble model 76.4(final submission) 78.1 80.0 Single model

  • 77.1

79.0 Year 2016 2017 2018 mmAP 60.5 72.1 76.4(ours)

Results on COCO test challenge recent years Results of our method

slide-5
SLIDE 5

Overview

  • Pipeline
  • Proposed Method
  • Experiments
  • Summary
slide-6
SLIDE 6

Overview

▪

Pipeline

slide-7
SLIDE 7

Pipeline

Single Person Pose Estimation Network

MegDet[1] crop

[1] Megdet: A large mini-batch object detector：C Peng, eta. (CVPR2018)

Human AP 66.0

slide-8
SLIDE 8

Overview

▪

Pipeline

▪

Proposed Method

slide-9
SLIDE 9

Proposed Methods

Cascade Pyramid Network[1]

[1] Cascaded pyramid network for multi-person pose estimation: Y Chen eta. (CVPR2018)

slide-10
SLIDE 10

Proposed Methods

Cascade Pyramid Network[1]

[1] Cascaded pyramid network for multi-person pose estimation: Y Chen eta. (CVPR2018)

head backbone

slide-11
SLIDE 11

Proposed Methods

Cascade Pyramid Network[1]

[1] Cascaded pyramid network for multi-person pose estimation: Y Chen eta. (CVPR2018)

head backbone

slide-12
SLIDE 12

Proposed Methods

BackBone of CPN

slide-13
SLIDE 13

Proposed Methods

Cascade Backbone Network

slide-14
SLIDE 14

Proposed Methods

Cascade Backbone Network with skip connection

slide-15
SLIDE 15

Proposed Methods

Cascade Backbone Network • Online Hard Keypoints Ming (OHKM) supervision

  • Coarse to fine supervision
slide-16
SLIDE 16

Overview

  • Pipeline
  • Proposed Method
  • Experiments
slide-17
SLIDE 17

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192

slide-18
SLIDE 18

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192

slide-19
SLIDE 19

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192 2res50+up-skip 73.4 256x192

slide-20
SLIDE 20

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192 2res50+up-skip 73.4 256x192 2res50+up-skip+down-skip 74.8 256x192

slide-21
SLIDE 21

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192 2res50+up-skip 73.4 256x192 2res50+up-skip+down-skip 74.8 256x192 2res50+skip+coarse to fine loss 75.0 256x192

slide-22
SLIDE 22

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192 2res50+up-skip 73.4 256x192 2res50+up-skip+down-skip 74.8 256x192 2res50+skip+coarse to fine loss 75.0 256x192 2res50+skip+coarse to fine loss 75.8 384x288

slide-23
SLIDE 23

Experiments

  • Ablation experiments

Baseline(backbone res50) 72.3 256x192 2res50 72.8 256x192 2res50+up-skip 73.4 256x192 2res50+up-skip+down-skip 74.8 256x192 2res50+skip+coarse to fine loss 75.0 256x192 2res50+skip+coarse to fine loss 75.8 384x288 3res50+skip+coarse to fine loss 77.5 384x288 4res50+skip+coarse to fine loss 79.0 384x288 Ensemble 80.0

slide-24
SLIDE 24

Experiments

▪ Data augmentation

  • Half body augmentation(+0.4 AP)

30% probability if more than 8 labeled keypoints

slide-25
SLIDE 25

Experiments

▪ Data augmentation

  • Half body augmentation(+0.4 AP)
  • Others(+0.4 AP)
  • Random scales(0.7 ~ 1.35)
  • Rotation (-45º ~ 45º)
  • Random flip
  • Ensemble(+1.0 AP)
  • External data(+0.7 AP)
  • Private data with 10W boxes
slide-26
SLIDE 26

Results

Prediction Ground truth

slide-27
SLIDE 27

Results

Prediction Ground truth

slide-28
SLIDE 28

Results

Prediction Ground truth

slide-29
SLIDE 29

Results

Prediction Ground truth

slide-30
SLIDE 30

Overview

  • Pipeline
  • Proposed Method
  • Experiments
  • Summary
slide-31
SLIDE 31

Summary

  • Cascade backbone
  • OHKM and coarse to fine supervision
  • Skip connections cross stages.
slide-32
SLIDE 32

Looking for Intern, Researcher, Research Engineer career@megvii.com yugang@megvii.com