1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn...

16
1 Accurate Object Detection with Joint Classification-Regression Random Forests Presenter ByungIn Yoo CS688/WST665

Transcript of 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn...

Page 1: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

1

Accurate Object Detection with Joint Classification-

Regression Random Forests

Presenter ByungIn Yoo

CS688/WST665

Page 2: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

2

Contents

● Introduction

● Motivation

● Main Idea

● Details

● Experiments

● Conclusion

Page 3: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

3

Introduction

● Object detection is one of the most important task for objects recognition as well as images searching.

● Object Detection + Bounding-box Regression (Localization)in a sliding window approach with a single model (Random Forest).

Classification: Car

Regression: 1(width):0.6(height)

Page 4: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

4

● Problems● Low-accurate Bounding-box Localization

● Low-accurate Label Classification

Motivation

89.3% overlap(Proposed method)

59.6% overlap(Previous methods)

Ground-truth

How we improve performances of localization and classification simultaneously?

Page 5: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

5

Main Idea

● Joint Classification-Regression Random Forest (JCRF)

Classification: Predict object probability

Regression: Estimate bounding-box aspect ratio

● What things are novel?

More accurate object detection and localization method

in a single model!

Training data: Car region, Background: Aspect ratio of the region

……

Tree1 Treet

Obj./Back. Aspectratio

x

0.8Obj./Back. Aspect

ratio

0.7

x

Random Forest (JCRF)

Testing Result: Car region (x,y): Aspect ratio of the region 0.75 (width-height)

Training Testing

Page 6: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

6

● Training Data =

image = {height x width x 10 feature channels}

Details – Object Detection Model

Label = {background, object}

Actual height (width is normalized to 100 pixels)

Blue boxes ( h x w ) show positive training data(Object).Zi show actual height of objects.

Negative data(Background)

Page 7: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

7

Details – Training JCRF (1/3)

● What is Random Forest? : Ensemble of multiple decision trees

● Split Node: Find and store a best splitting parameter.

● Leaf Node: Store class probabilities and an aspect ratio.

……

Tree1 Treet

Obj./Back. Aspectratio

x

0.8Obj./Back. Aspect

ratio

0.7

xSplit Node

Leaf Node

Simply averaging class probabilities and aspect ratio from all trees!

Page 8: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

8

Details – Training JCRF (2/3)

● Two split node types are employed.

● Binary Classification Node ( Object or Background? )

● Regression Node ( How long is the object height? )

● Randomly decided which types are being optimized in an each split node.

● Classification Split nodes

● Objective: Find a good splitting rule which minimize the Entropy of the classes between left and right dataset.

● Regression Split nodes

● Objective: Find a good splitting rule which minimize the height variance of the between left and right dataset.

Page 9: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

9

● Learning procedure for each split node

1) Randomly select parameter

2) Split data into two sets by the

3) Evaluate Information Gain ( Is the minimize I(·)? )

4) Repeat 1)~3) until finding best parameter. Store the parameter.

Details – Training JCRF (3/3)

𝒄=𝟏 𝒄=𝟐 10

Location1 Location2 Threshold Channel𝑶𝒃𝒋 .

Back

Training Data

What is the best parameter to split?

Splitting functionPixel value 1of channel c

Pixel value 2of channel c

threshold

Shannon Entropy

Location1

Location2

Page 10: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

10

Details – Testing JCRF (1/2)

● For detecting objects in test images, a standard sliding window W is utilized.

● Detection score s of a given image x in a window W

● Object height z of a given image x in a window W

● The resulting detection D of a window W

k is the scale of the detection

x,y: location of W w: width of W z: height of W

s: Object detection score of W

FR: Regression Function of a JCRFT: Number of Trees in a JCRF

FC: Classification Function of a JCRFT: Number of Trees in a JCRF

Page 11: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

11

Details – Testing JCRF (2/2)

● Early stopping is utilized to boost a testing speed.

tree 1 tree T

x x

tree t

x

Early Stopping Criterion:

Testing Progress

Detection threshold

Current summation of scores Upper bound of remaining scores

Page 12: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

12

Experiments (1/3)

● Evaluation Criterion: Pascal overlap IoU (Intersection over Union)

Detected Region Ground-truth

:True

:False

Page 13: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

13

Experiments (2/3)

● Precision-recall curve for the bounding-box accuracy.

Proposed methods shows best performance.

[ TUDpedestrian Dataset ]

Page 14: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

14

Experiments (3/3)

● Tightening the Pascal Overlap Criterion

Proposed methods shows best performance.

[ ETHZcars Dataset ]

Page 15: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

15

Conclusion

● Random forest based object detection and predicting aspect-ratio method is proposed.

● Joint Classification-Regression Forest exploits class labels as well as actual heights during both training and testing.

● Proposed detection model recognize more accurate object regions than related state-of-the art approaches.

Page 16: 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665.

16

Appendix – More Experiments

● Increasing performance until number of trees=100.

● Parameter selection Map for Splitting nodes

Saturation!

Classification only Classification + Regression

Diverse locations! Separate Different Views