Filtered Channel Features for Pedestrian Detection Rodrigo - - PowerPoint PPT Presentation

▶
filtered channel features for
SMART_READER_LITE
LIVE PREVIEW

Filtered Channel Features for Pedestrian Detection Rodrigo - - PowerPoint PPT Presentation

Jan 23, 2024 247 likes •579 views

Filtered Channel Features for Pedestrian Detection Rodrigo Benenson Shanshan Zhang Bernt Schiele Cooperation Prof. Luc Van Gool (ETH & KU Leuven) Prof. Bernt Schiele Shanshan Zhang | Valse Webinar 2 Outline Background

slide-1
SLIDE 1

Rodrigo Benenson

Filtered Channel Features for Pedestrian Detection

Bernt Schiele

Shanshan Zhang

slide-2
SLIDE 2

Shanshan Zhang | Valse Webinar

2

Cooperation

  • Prof. Luc Van Gool (ETH & KU Leuven)
  • Prof. Bernt Schiele
slide-3
SLIDE 3

Shanshan Zhang | Valse Webinar

3

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-4
SLIDE 4

Shanshan Zhang | Valse Webinar

4

Google driverless car

Sebastian Thrun (Google & Stanford)

slide-5
SLIDE 5

Shanshan Zhang | Valse Webinar

5

Sensors

Camera

slide-6
SLIDE 6

Shanshan Zhang | Valse Webinar

6

slide-7
SLIDE 7

Shanshan Zhang | Valse Webinar

7

Benchmarks

  • Caltech

http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/

  • KITTI

http://www.cvlibs.net/datasets/kitti/

slide-8
SLIDE 8

Shanshan Zhang | Valse Webinar

8

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-9
SLIDE 9

Shanshan Zhang | Valse Webinar

9

ACF

  • P. Dollar, R. Appel, S. Belongie, and P. Perona. Fast Feature Pyramids for Object Detection. PAMI 2014.
  • P. Dollar, Z. Tu, P. Perona and S. Belongie. Integral Channel Features. BMVC 2009.

ICF

slide-10
SLIDE 10

Shanshan Zhang | Valse Webinar

10

InformedHaar

  • S. Zhang, C. Bauckhage, A. B. Cremers. Informed Haar-like Features Improve Pedestrian Detection. CVPR 2014.

Pedestrian shape model

slide-11
SLIDE 11

Shanshan Zhang | Valse Webinar

11

InformedHaar

 Which body part is more

discriminative?

 Robustness against

variations from

 Occlusion  Viewpoints

Accumulative weight map

slide-12
SLIDE 12

Shanshan Zhang | Valse Webinar

12

LDCF

  • W. Nam, P. Dollar, and J. H. Han. Local Decorrelation for Improved Pedestrian Detection. NIPS 2014.

Learned decorrelation filters Original and decorrelated channels averaged over positive samples

slide-13
SLIDE 13

Shanshan Zhang | Valse Webinar

13

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-14
SLIDE 14

Shanshan Zhang | Valse Webinar

14

Unified framework

slide-15
SLIDE 15

Shanshan Zhang | Valse Webinar

15

Filter bank families

Which one is the best ?

SquaresChntrs

  • R. Benenson, M. Mathias, T. Tuytelaars and L. Van Gool. Seeking the Strongest Rigid Detector. CVPR 2013.

InformedFilters

  • S. Zhang, C. Bauckhage, A. B. Cremers. Informed Haar-like Features Improve Pedestrian Detection. CVPR 2014.

LDCF

  • W. Nam, P. Dollar, and J. H. Han. Local Decorrelation for Improved Pedestrian Detection. NIPS 2014.
slide-16
SLIDE 16

Shanshan Zhang | Valse Webinar

16

InformedFilters

 Filter contents: informed  Filter positions: everywhere  How many identical filters?

 Redundancy removal  212(209+3) filters by 4x3

slide-17
SLIDE 17

Shanshan Zhang | Valse Webinar

17

Checkerboards

 Blurring filters  Horizontal filters  Vertical filters  How many filters?

 7 filters by 2x2  25 filters by 3x3  39 filters by 4x3  61 filters by 4x4

slide-18
SLIDE 18

Shanshan Zhang | Valse Webinar

18

PCA filters

 Background (4 filters)  Foreground (4 filters)

slide-19
SLIDE 19

Shanshan Zhang | Valse Webinar

19

RandomFilters

 Random

 15  50  ...

 Selected from random

 Top 15 / 50xN  Top 50 / 50xN  ...

slide-20
SLIDE 20

Shanshan Zhang | Valse Webinar

20

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-21
SLIDE 21

Shanshan Zhang | Valse Webinar

21

How many filters?

slide-22
SLIDE 22

Shanshan Zhang | Valse Webinar

22

More data & deeper trees

slide-23
SLIDE 23

Shanshan Zhang | Valse Webinar

23

Ingradients to improve

Which ingradient is the most important ?

slide-24
SLIDE 24

Shanshan Zhang | Valse Webinar

24

Comparison of improvements

slide-25
SLIDE 25

Shanshan Zhang | Valse Webinar

25

Add-ons

  • Context: 2Ped
  • ~5.0pp improvement (Ouyang etal. CVPR 2013)
  • ~2.8pp improvement (Benenson etal. ECCV Workshop 2014)
  • <0.5pp improvement (Checkerboards)
  • W. Ouyang and X. Wang. Single-pedestrian Detection Aided by Multi-pedestrian Detection. CVPR 2013.
slide-26
SLIDE 26

Shanshan Zhang | Valse Webinar

26

Add-ons

  • Flow: SDt
  • ~7pp improvement (ACF)
  • ~5pp improvement (Benenson etal. ECCV Workshop 2014)
  • 1.4pp improvement (Checkerboards)
  • D. Park, C. L. Zitnick, D. Ramanan, and P. Dollar. Exploring Weak Stabilization for Motion Feature Extraction.

CVPR 2013.

slide-27
SLIDE 27

Shanshan Zhang | Valse Webinar

27

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-28
SLIDE 28

Shanshan Zhang | Valse Webinar

28

Test set results

slide-29
SLIDE 29

Shanshan Zhang | Valse Webinar

29

Test set results

Caltech test set

slide-30
SLIDE 30

Shanshan Zhang | Valse Webinar

30

Outline

 Background  Baseline detectors

 ACF  InformedHaar  LDCF

 Unified framework  Experimental setup  Test set results  Take-away messages

slide-31
SLIDE 31

Shanshan Zhang | Valse Webinar

31

Take-away messages

  • Filtered channel features are great!
  • There is no flagrant difference between different filter types.
  • More training data and deeper trees help a lot.
  • The improvements from current context models and optical flow

features become smaller as the baseline detector grows stronger.

slide-32
SLIDE 32

Shanshan Zhang | Valse Webinar

32

Thank you for your attention!

Q & A