Thesis report and full details: https://imatge.upc.edu/web/publications/contextless-object-recognition-shape-enriched-sift-and-bags-features
Author: Marcel Tella
Advisors: Xavier Giró-i-Nieto (UPC) and Matthias Zeppelzauer (TU Wien)
Degree: Telecommunications Engineering (5 years) at Telecom BCN-ETSETB (UPC)
Abstract:
Currently, there are highly competitive results in the field of object recognition based on the aggregation of point-based features. The aggregation process, typically with an average or max-pooling of the features generates a single vector that represents the image or region that contains the object.
The aggregated point-based features typically describe the texture around the points with descriptors such as SIFT. These descriptors present limitations for wired and textureless objects. A possible solution is the addition of shape-based information. Shape descriptors have been previously used to encode shape information and thus, recognise those types of objects. But generally an alignment step is required in order to match every point from one shape to other ones. The computational cost of the similarity assessment is high.
We purpose to enrich location and texture-based features with shape-based ones. Two main architectures are explored: On the one side, to enrich the SIFT descriptors with shape information before they are aggregated. On the other side, to create the standard Bag of Words histogram and concatenate a shape histogram, classifying them as a single vector.
We evaluate the proposed techniques and the novel features on the Caltech-101 dataset.
Results show that shape features increase the final performance. Our extension of the Bag of Words with a shape-based histogram(BoW+S) results in better performance. However, for a high number of shape features, BoW+S and enriched SIFT architectures tend to converge.
Contextless Object Recognition with Shape-enriched SIFT and Bags of Features
1. Contextless Object Recognition
with Shape-enriched SIFT and
Bags of Features
Marcel Tella Amo
Directed by Dr. Matthias Zeppelzauer (TU Wien)
Codirected by Dr. Xavier Giró-i-Nieto (UPC)
2. Motivation
2
Object Recognition and Classification
Categories
• Ball
• Airplane
• Chair
• Beaver
• …
Ball Airplane Chair
Shape
Information
Texture
information
5. Requirements State of the Art Design Results
Design shape features that can be used in an
aggregated framework, like Bag of Words with
no need of matching or alignment.
5
Take a
successful method :
Shape
Information
SIFT
6. Requirements State of the Art Design Results
Analyse the implication of the vocabulary size
with respect to the size of the shape features.
SIFT
6
Shape
7. The proposed features should be at least scale,
rotation and translation invariant. If it is
possible, flip invariant as well.
7
Requirements State of the Art Design Results
8. Need for Segmentation to codify the shape
Study the limitations of shape coding when using a state of the art
segmentation.
Manual annotations vs Automatic Segmentation
8
Requirements State of the Art Design Results
10. Requirements State of the Art Design Results
Object Candidates algorithms
Multiscale Combinatorial Grouping (MCG)
10
Ranking
Object Plausibility
Arbelaez, P., Pont-Tuset, J., Barron, J. T., Marques, F., Malik, J. (2014).
Multiscale Combinatorial Grouping. CVPR.
High
Low
11. Requirements State of the Art Design Results
Shape Context
11
G. Mori, S. Belongie, and J. Malik. Ecient shape
matching using shape
contexts. PAMI, 27(11), 2005.
12. Requirements State of the Art Design Results
Interest point descriptors:
SIFT descriptor
Simplified example
Typically 4x4 divisions * 8 bins/hist = 128 features
dense SIFT
sparse SIFT
12
David G Lowe, Distinctive image features from scale-invariant keypoints, International journal of
computer vision 60 (2004), no. 2, 91{110.
13. Requirements State of the Art Design Results
Enrichment of SIFT
Extra features : Absolute spatial location (X,Y) or angle and distance
Rene Grzeszick, Leonard Rothacker, and Gernot A. Fink, "Bag-of-features representations using spatial visual vocabularies
for object classication,“ in IEEE Intl. Conf. on Image Processing, Melbourne, Australia, 2013
Extra features : Relative position + aspect ratio + scale ratio + Color Space
Carreira, J., Caseiro, R., Batista, J., & Sminchisescu, C. (2012). Semantic segmentation with second-order pooling. In
Computer Vision{ECCV 2012} (pp. 430-443). Springer Berlin Heidelberg.
13
128-dimensional SIFT descriptor Extra features
14. Bag of Words
14
Requirements State of the Art Design Results
15. Requirements State of the Art Design Results
Bags of Words - Pipeline
15
Get
Descriptors
Clustering
(K-means)
Create
histograms
Train Model
(SVM)
Image
Create
histogram
Evaluate
(SVM)
18. Main principle: Combination of dense SIFT and Object Candidates
18
Requirements State of the Art Design Results
19. Requirements State of the Art Design Results
Distance to the nearest border (DNB)
Logarithmic distance to the nearest border (LDNB)
Less influence of big distances
19
Carreira, J., Caseiro, R., Batista, J., & Sminchisescu, C. (2012). Semantic segmentation with second-order
pooling. In Computer Vision-ECCV 2012 (pp. 430-443). Springer Berlin Heidelberg.
20. Distance and Angle to the nearest border (DANB)
Problem: Really similar in 2D but very different values.
Solution: Codify them in two separated features.
20
Requirements State of the Art Design Results
22. Distance to the center (DC)
22
Requirements State of the Art Design Results
23. η - Angular Scan (ηAS)
WINNER!
23
Requirements State of the Art Design Results
24. Shape Context from a dense SIFT (DSC)
Note: It crosses the contour of the region like Shape Context.
ηAS does not!
24
Requirements State of the Art Design Results
25. Requirements State of the Art Design Results
Rotation Invariant Region Quantization (RIRQ)
Main idea: Get spatial information.
Easily extensible to a pyramid!
25
Lazebnik, S., Schmid, C., & Ponce, J. (2006). 2006 IEEE Computer Society Conference on (Vol. 2, pp.
2169-2178). IEEE.
26. Achieving flip invariance (RIRQ)
1
2
4 3
1
2 3
4
2
4 1
3 2
3
4
1
4 2 2 4
SORT SORT
2 4
26
Requirements State of the Art Design Results
27. Where do we integrate our features?
Two main Architectures
Enriched SIFT (eSIFT)
SIFT Shape features
Visual Vocabulary
Bag of eSIFT visual words
BoW+Shape
SIFT
Visual Vocabulary
Bag of Words Shape histogram
27
Requirements State of the Art Design Results
28. BoW+Shape Creation of the shape histograms
SIFT
Accumulation of features
Visual Vocabulary
Bag of Words Shape histogram
1
1. Accumulate the
same feature for all
points .
2. Create a
histogram of X bins
for that feature.
1
2
2
3. Concatenate
histograms to create
the final one.
Example: 8-Angular Scan
8 distances (different angles)
# SIFT keypoints
28
Requirements State of the Art Design Results
30. Requirements State of the Art Design Results
The dataset: Caltech-101
30
•Well recognized dataset
• 101 Different Categories of images
• Ground truth annotations available
• From 40 to 800 images per category.
31. Requirements State of the Art Design Results
Metrics: Accuracy (%)
31
Correct Classifications
Correct + Incorrect Classifications
32. Requirements State of the Art Design Results
Experiments setup
32
• 30 images per category in train and 30-50 in test.
• 101 Categories + Background category.
• Different Vocabulary sizes in the X axis.
• Accuracy(%) in the Y axis:
•Experiments and analysis:
• eSIFT
• BoW+S
• eSIFT vs BoW+S
• Performance acheived
• Comparison between adding features before or after quantization
• Number of bins per histogram
• Ground truth vs MCG Object Canditates
• Context vs Shape
35. Requirements State of the Art Design Results
Performance achieved
35
Conclusion
With Angular Scan, there is an increase of performance
from 16% to around 41%.
36. Requirements State of the Art Design Results
Comparison between adding features
after and before
Conclusion
In Angular Scan, if the number of shape features is high,
both architectures tend to converge. 36
37. Requirements State of the Art Design Results
Number of bins per histogram
Conclusion
In Angular Scan, 8 bins is the value that gives the best
performance. 37
38. Requirements State of the Art Design Results
Ground truth vs MCG Object Candidates
Conclusion 1
2
Higher vocabulary values lead to a more robust
approach in terms of segmentation errors.
Shape-based methods are more sensible to
segmentation errors than texture-based. 38
39. Requirements State of the Art Design Results
Context gain vs Shape gain
Conclusion
Object
Context
It gives better performance to codify the shape
than the context of the image. 39
40. FutureWork
Comparison betwen our work and
Second Order Pooling
PhD thesis of Carles Ventura
Carreira, J., Caseiro, R., Batista, J., & Sminchisescu, C. (2012). Semantic segmentation with second-order
pooling. In Computer Vision-ECCV 2012 (pp. 430-443). Springer Berlin Heidelberg.
40
42. Conclusions
1. Increase of performance from 16% to around 41%
2. In Angular Scan, if the number of shape features is high, both
architectures tend to converge.
3. In Angular Scan, 8 bins is the value that gives the best performance.
4. Higher vocabulary values lead to a more robust approach in terms of
segmentation errors.
5. Shape-based methods are more sensible to segmentation errors than
texture-based.
6. It gives better performance to codify the shape than the context of the
image.
Thank you!
Questions? 42