Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and...

37
Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1

description

Introduction  Object detector : Top-down approaches 3

Transcript of Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and...

Page 1: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

1

Object Recognition as Ranking Holistic Figure-Ground

Hypotheses

Fuxin Li and Joao Carreira and Cristian Sminchisescu

Page 2: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

2

Outline Introduction Method Overview Segment Categorization Segment Post-Processing Experiment Conclusion

Page 3: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

3

Introduction Object detector : Top-down approaches

Page 4: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

4

Introduction Semantic segmentation results

produced by our algorithm

Page 5: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

5

Introduction Ideally segment1. Can model entire object

2. At least sufficiently distinct parts of them

Page 6: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

6

Page 7: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

7

Introduction Constrained Parametric Min Cuts

algorithm (CPMC) [6]

Page 8: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

8

Method Overview

This paper focus

CPMC

Number of segment

Page 9: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

9

Method Overview Recognition framework1. Segment categorization

2. Segment post-processing

Page 10: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

10

Segment categorization1. Scoring function2. Sort3. Combine high-rank segment

Page 11: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

11

Page 12: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

12

Segment post-processing

COW

Page 13: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

13

Segment Categorization Multiple Segment and Features Learning Scoring Functions with

Regression Learning the Kernel Hyperparameters Compare with Structural SVM

Page 14: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

14

Multiple Segment and Features

Model object appearance : 1. Extracted four bag of words of SIFT 2. Two on foreground 3. Two on Background, aim to improve

recognition

Page 15: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

15

Multiple Segment and Features

Encode shape information :1. A bag of word of local sharp

contexts [2] : measure similarity between shapes

2. Three pyramid HOGs [5] : classifying images by the object

categories they contain

Page 16: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

16

Multiple Segment and Features

Chi-square kernel :

Computed from each histogram feature and use a weighted sum of such kernel for regression

Page 17: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

17

Learning Scoring Functions with Regression

: Image I with ground truth segments

: Segmentation algorithm provides a set

of segment : Denote the K object

categories : Indicator function

Page 18: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

18

Learning Scoring Functions with Regression

Quality function :

Measure overlap with all denote

the value for and is the maximal overlap with ground truth segments belonging to , and do not appear

Page 19: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

19

Learning Scoring Functions with Regression

Learn the function for each :

use nonlinear SVR(Support Vector Regression) to regress against ,the features extracted from

Page 20: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

20

Learning Scoring Functions with Regression

Use kernel trick : : support vector from training set : obtained by the SVR optimizer : maximal score of the

segment : final class of the

segment

Page 21: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

21

Learning the Kernel Hyperparameters

Fundamental equation (3) is infeasible to estimate all kernel hyperparameter via grid search

Use subset of data comprised segments that best overlap each ground truth segment

Page 22: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

22

Compare with Structural SVM The structural SVM(in [3])

formulation for sliding window prediction is :

Page 23: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

23

Connections with Structural SVM

Our algorithm VS Structural SVM Structural SVM score the bounding box

and Our algorithm score the segment

Important advantage1. Guarantee the highest rank for the

ground truth2. Correct ranking for all segment

Page 24: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

24

Segment Post-Processing Simple decision rule : avoid the post-

processing and direct choose the segment , cannot detect multiple objects

Our methodology : weighted consolidation of segment and sequential interpretation strategy

Page 25: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

25

Segment Post-Processing

Page 26: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

26

Segment Post-Processing To decide which segments to

combine

Consider segment with intersection

> 0.75 for combination

Page 27: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

27

Segment Post-Processing1. Highest-scoring segment as seed2. Group segments that intersect it3. Generated a final mask 4. Proceed with the next higher rank

segment5. Choose segment that are not

overlapping with 3

Page 28: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

28

Segment Post-Processing Generate the score for the pixels in

the mask by (9), only pixels with score > 0.65 are displayed in the mask.

Page 29: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

29

Experiments Classification : Caltech-101 Detection : ETHZ Shape classes Segmentation : VOC 2009

Page 30: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

30

Classification

Page 31: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

31

Classification

Page 32: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

32

Detection

Page 33: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

33

Detection

Page 34: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

34

Detection

Page 35: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

35

OWT-UCM Masks

Page 36: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

36

Segmentation

Bounding box

Page 37: Object Recognition as Ranking Holistic Figure-Ground Hypotheses Fuxin Li and Joao Carreira and Cristian Sminchisescu 1.

37

Conclusion CPMC Categorization Post-

processing Achieve good performance Future work : improve the scalability

to be able to process hundreds of thousands of image