Challenges 2018 Megvii (Face++) Team lizeming@megvii.com I. COCO1 8 - - PowerPoint PPT Presentation

▶
challenges 2018
SMART_READER_LITE
LIVE PREVIEW

Challenges 2018 Megvii (Face++) Team lizeming@megvii.com I. COCO1 8 - - PowerPoint PPT Presentation

Oct 22, 2022 313 likes •691 views

MSCOCO Instance Segmentation Challenges 2018 Megvii (Face++) Team lizeming@megvii.com I. COCO1 8 Instance Seg Zeming LI Jian SUN Yueqing ZHUANG Xiangyu ZHANG Gang YU Overview Improvements The results is obtained on test-dev Mask mmAP

slide-1
SLIDE 1

MSCOCO Instance Segmentation Challenges 2018

Megvii (Face++) Team lizeming@megvii.com

slide-2
SLIDE 2
  • I. COCO’18 Instance Seg

Zeming LI Yueqing ZHUANG Gang YU Jian SUN Xiangyu ZHANG

slide-3
SLIDE 3

Overview

37.4 41.6 52.6 56.0 35 40 45 50 55 60

2015 2016 2017(Megvii) Ours

Detector mmAP

28.4 37.6 46.7 48.8

25 30 35 40 45 50 55

2015 2016 2017 Ours

Mask mmAP

Object Detector 3.4% improvement

The results is obtained on test-dev

Instance Segmentation 2.1% improvement

Improvements

slide-4
SLIDE 4

Outline

1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4) Results

slide-5
SLIDE 5

Outline

1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

slide-6
SLIDE 6

FPN Original Mask Head

Instance Seg mmAP Det mmAP Original Paper(detectron 1x) 33.6

  • Our Re-implement

34.4 37.0

Mask RCNN Baseline

slide-7
SLIDE 7

Location Sensitive Header

Overall Architecture Comparison

slide-8
SLIDE 8

1) Location Sensitive Detector

name Mask AP Bbox AP Improvement Baseline 34.4 37.0

  • + Local Sensitive Detector

35.4 38.7 + 1.0 / +1.7

Location Sensitive Header

slide-9
SLIDE 9

2) Multi-Scale RoI

name Mask AP Bbox AP Improvement Baseline 34.4 37.0

  • + Local Sensitive Detector

35.6 38.7 + 1.0 / +1.7 + Multi-Scale RoI 35.8 38.9 + 0.2 / +0.2

Location Sensitive Header

slide-10
SLIDE 10

3) Heavier Header

name Mask AP Bbox AP Improvement Baseline 34.4 37.0

  • Heavier Header

35.3 36.8 + 0.9 / -0.2

Location Sensitive Header

slide-11
SLIDE 11

4) Mask Edge Loss

name Mask AP Bbox AP Improvement Baseline 34.4 37.0

  • Mask Edge Loss

35.0 37.0 + 0.6 / +0.0

Location Sensitive Header

slide-12
SLIDE 12

4) Mask Edge Loss

Location Sensitive Header

Sigmoid Cross Entropy

slide-13
SLIDE 13

Location Sensitive Header

Review of overall Architecture

Location Sensitive Header: 1) Location Sensitive Detector 2) Multi-Scale RoI 3) Heavier Header 4) Mask Edge Loss

slide-14
SLIDE 14

Overall Performance in Small and Large Model

BackBone Header Mask AP Bbox AP Improvement ResNet50 Baseline 34.4 37.0

  • Location Sensitive Header

37.0 39.3 + 2.6 / + 2.0 ShuffleV2-GAP Baseline 40.3 45.0

  • Location Sensitive Header

42.3 46.5 +2.0/+1.5

Location Sensitive Header

We will introduce backbone in next slides

slide-15
SLIDE 15

Outline

1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

slide-16
SLIDE 16

Backbone Improvement

  • 1. Channel Information Flow

Ma N, Zhang X, Zheng H T, et al. ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design[J]. 2018.

slide-17
SLIDE 17
  • 2. Add Global Information

name Mask AP Bbox AP Improvement Baseline 34.4 37.0

  • +GAP

35.1 37.7 +0.7/+ 0.7

Backbone Improvement

slide-18
SLIDE 18

Outline

1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

slide-19
SLIDE 19

Two-Pass Pipeline

slide-20
SLIDE 20

Outline

1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

slide-21
SLIDE 21

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 2x Means 2x training setting used in Detectron Trained On Megvii’s Megbrain

slide-22
SLIDE 22

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 + Location Sensitive Header 42.3 46.5 +2.0 /+1.5 Trained On Megvii’s Megbrain

slide-23
SLIDE 23

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU + Multi Scale Training + BN training 44.5 49.3 +2.2/ 2.8 Trained On Megvii’s Megbrain

slide-24
SLIDE 24

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU + Multi Scale Training + BN training 44.5 49.3 +2.2/ 2.8 + Improve on Dets 47.6 55.4 +3.1/ 6.1 Trained On Megvii’s Megbrain

slide-25
SLIDE 25

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU + Multi Scale Training + BN training 44.5 49.3 +2.2/ 2.8 + Improve on Dets 47.6 55.4 +3.1/ 6.1 + Seg Multi-scale Testing 48.1 55.4 +0.5/0.0 Trained On Megvii’s Megbrain

slide-26
SLIDE 26

Results

name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3

  • ShuffleV2 (1batch)

40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU + Multi Scale Training + BN training 44.5 49.3 +2.2/ 2.8 + Improve on Dets 47.6 55.4 +3.1/ 6.1 + Seg Multi-scale Testing 48.1/ 48.8(dev) 55.4/ 56.0(dev) +0.5/0.0

Instance Segmentation is obtained by single instance segmentation model

Trained On Megvii’s Megbrain

slide-27
SLIDE 27

Results

name Bbox AP(val) Improvement Baseline 49.3

  • +Soft-Nms

49.8 +0.5 +Multi-scale Testing 51.6 +1.8 +Ensemble 53.6 +2.0 add an additional model for ensemble: +with cascade R-CNN +external COCO++ 11W data 55.4 +1.8 Trained On Megvii’s Megbrain

slide-28
SLIDE 28

Results

slide-29
SLIDE 29

Visualization Comparison

Our baseline Location Sensitive Header

Refine Location Error

slide-30
SLIDE 30

Visualization Comparison

Our Baseline Location Sensitive Header

Refine Location Error

slide-31
SLIDE 31

Visualization Comparison

Our Baseline Location Sensitive Header

Refine Location Error

slide-32
SLIDE 32

Visualization Comparison

Our Baseline Location Sensitive Header

slide-33
SLIDE 33

Visualization Comparison

Our Baseline Location Sensitive Header

slide-34
SLIDE 34

Visualization

Detector Results Mask Results

slide-35
SLIDE 35

Visualization

Detector Results Mask Results

slide-36
SLIDE 36

Summary & thanks

  • 1. Location Sensitive Header
  • 2. Backbone Improvement
  • 3. Pipeline Optimization

Other Improvements:

  • 1. Multi-Scale Training
  • 2. Large Batch (MegDet : [C. Peng, CVPR’ 18])
  • 3. Multi-Scale and Flip Testing
  • 4. Ensemble (only for Detection)
slide-37
SLIDE 37

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