Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Learning Object Detectors with Weak Supervision
Kun He
Committee members:
Prof. Stan Sclaroff
Prof. Margrit Betke
Prof. Pedro Felzenszwalb
Problem: object detection
09/24/2014Learning Object Detectors with Weak Supervision
2
Source: The PASCAL Visual Object Classes Challenge 2007
Supervised learning pipeline
09/24/2014Learning Object Detectors with Weak Supervision
3
• Image credit: Sudheendra Vijayanarasimhan
What about annotations?
• Example: Microsoft COCO (Lin et al ECCV’14)
09/24/2014Learning Object Detectors with Weak Supervision
4
What about annotations?
Image credit: Tsung-Yi Lin
09/24/2014Learning Object Detectors with Weak Supervision
5
What about annotations?
09/24/2014Learning Object Detectors with Weak Supervision
6
Example taken from Microsoft COCO dataset http://mscoco.org/explore/?id=79387
What about annotations?
09/24/2014Learning Object Detectors with Weak Supervision
7
Example taken from Microsoft COCO dataset http://mscoco.org/explore/?id=79387
Relaxing annotation requirements
• Annotation process: laborious & error-prone
• Learn directly from the images! (weak supervision)
09/24/2014Learning Object Detectors with Weak Supervision
8
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
9
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
10
Generative part-based models
• Detect sparse features → fit part-based model → determine (non-)existence of object
• Rob Fergus, Pietro Perona and Andrew Zisserman, CVPR’03
09/24/2014Learning Object Detectors with Weak Supervision
11
Generative models (Fergus et al CVPR’03)
09/24/2014Learning Object Detectors with Weak Supervision
12
• Likelihood ratio test
• Likelihood: product of Gaussians• Features: location X, scale S, appearance A
• h : hypothesis (part-based object configuration)
Foreground model
Background model
Generative models (Fergus et al CVPR’03)
• Learning: maximum likelihood via EM• E-step: expectation wrt. ℎ
• M-step: update Gaussian parameters
09/24/2014Learning Object Detectors with Weak Supervision
13
Generative models (Fergus et al CVPR’03)
• Learning: maximum likelihood via EM• E-step: expectation wrt. ℎ
• M-step: update Gaussian parameters
09/24/2014Learning Object Detectors with Weak Supervision
14
𝑂(#𝑝𝑎𝑟𝑡𝑠#𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠)Typical: 630
Generative models (Fergus et al CVPR’03)
• Learned 6-part model for “face”
09/24/2014Learning Object Detectors with Weak Supervision
15
Generative models (Fergus et al CVPR’03)
• Face: single-Gaussian appearance model fails
09/24/2014Learning Object Detectors with Weak Supervision
16
Generative models (Fergus et al CVPR’03)
• Spotted cat: single-Gaussian shape model fails
09/24/2014Learning Object Detectors with Weak Supervision
17
Generative models: critiques
• GoodProbabilistic formulation
Models multiple factors
• Bad EM is slow
Limited modeling power
09/24/2014Learning Object Detectors with Weak Supervision
18
Generative models: critiques
• GoodProbabilistic formulation
Models multiple factors
• Bad EM is slow
Limited modeling power
• Discriminative models• Only model the decision boundary
• Usually perform better, eg. DPM (Felzenszwalb et al PAMI’10)
09/24/2014Learning Object Detectors with Weak Supervision
19
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
20
Multiple Instance Learning (MIL)
09/24/2014Learning Object Detectors with Weak Supervision
21
Image credit: Samarjit Das
Multiple Instance Learning (MIL)
• Images as bags
• Candidate generation• Segmentation [Galleguillos et al ECCV’08]
• Objectness [Alexe et al PAMI’12]
• Selective Search [Uijlings et al IJCV’13]
• EdgeBoxes [Zitnick & Dollar ECCV’14]
• ……
09/24/2014Learning Object Detectors with Weak Supervision
22
MIL for learning detectors
• Chicken-and-egg problem / latent variable model
Optimize(positive_instances, model_parameters)
• EM-like algorithms (MI-SVM, MI-CRF)• Impute latent variables
• Update model parameters
• Iterate
09/24/2014Learning Object Detectors with Weak Supervision
23
latent
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
24
SVM review
09/24/2014Learning Object Detectors with Weak Supervision
25
MI-SVM
09/24/2014Learning Object Detectors with Weak Supervision
26
MI-SVM
• “Witness”: identified positive instance within a positive bag
09/24/2014Learning Object Detectors with Weak Supervision
27
MI-SVM algorithm (Andrews et al NIPS’02)
09/24/2014Learning Object Detectors with Weak Supervision
28
1. Initialize
2. Update witnesses for positive bags•
3. Update model• solve fully-supervised SVM
4. Repeat
• Convergence: to local optimum
Progression of MI-SVM
• Source: R. Gokberk Cinbis, Jakob Verbeek and Cordelia Schmid, CVPR’14
09/24/2014Learning Object Detectors with Weak Supervision
29
MI-SVM: critiques
• GoodSimple optimization problem, solvers available
• Bad Sensitive to initialization
Witness update: no strong coupling between images
09/24/2014Learning Object Detectors with Weak Supervision
30
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
31
Conditional Random Fields (CRF) for MIL
• Enforce similarity between witnesses
09/24/2014Learning Object Detectors with Weak Supervision
32
MI-CRF (Deselaers et al IJCV’12)
• Pairwise CRF
09/24/2014Learning Object Detectors with Weak Supervision
33
“objectness”
similarity
MI-CRF (Deselaers et al IJCV’12)
• Pairwise CRF
• “Objectness”
• Ω: generic “objectness”
• Π: class-specific shape score
• Υ: class-specific appearance score
• Similarity
• Λ: shape similarity
• Γ: appearance similarity
09/24/2014Learning Object Detectors with Weak Supervision
34
MI-CRF: algorithm
09/24/2014Learning Object Detectors with Weak Supervision
35
Localize objects by
optimizing global energy
MI-CRF: results• Example detections
• Models learned by DPM (Felzenszwalb et al PAMI’10) vs. MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
36
MI-CRF: critiques
• GoodStrong coupling between images
• Bad High complexity (fully-connected CRF)
Limited #candidates per image (<=100)
09/24/2014Learning Object Detectors with Weak Supervision
37
Literature review outline
• Weber et al ECCV’00, Fergus et al CVPR’03, Crandall & Huttenlocher ECCV’06
Generative models
Discriminative: Multiple Instance Learning (MIL)
• Vijayanarasimhan & Grauman CVPR’08, Siva & Xiang ICCV’11, Cinbis et al CVPR’14, Song et al ICML’14 …MI-SVM
• Deselaers et al IJCV’12MI-CRF
09/24/2014Learning Object Detectors with Weak Supervision
38
Beyond MIL
• OPTIMOL: Li et al CVPR’07
• NEIL: Chen et al ICCV’13
Active learning
• Improving MI-SVM
Current research
09/24/2014Learning Object Detectors with Weak Supervision
39
Active learning
• Closing the loop
09/24/2014Learning Object Detectors with Weak Supervision
40
?
OPTIMOL (automatic Object Picture collecTion via Incremental MOdel Learning)
• Li-Jia Li, Gang Wang and Li Fei-Fei, CVPR’07
09/24/2014Learning Object Detectors with Weak Supervision
41
NEIL (Never-Ending Image Learner)
09/24/2014Learning Object Detectors with Weak Supervision
42
• Xinlei Chen, Abhinav Shrivastava and Abhinav Gupta, ICCV’13
Current research: improving MI-SVM
• MI-SVM (→ local optimum)1. Update witnesses independently
2. Update model parameters: solve SVM
• Idea: relax step 1 to
Still have convergence
Freedom to enforce desired properties
09/24/2014Learning Object Detectors with Weak Supervision
43
Current research: improving MI-SVM
• Enforcing similarity between witnesses
• Step t:
• Comparison: PASCAL VOC 2007, detection mAPcat cow dog
• MI-SVM 23.83, Ours 24.12
09/24/2014Learning Object Detectors with Weak Supervision
44
MI-SVM 34.8 43.7 22.2 10.4 7.8 36.2 22.0 20.6 11.1 21.4 28.7 38.0 19.6 23.7 19.8 35.4 9.8
Ours 38.9 42.4 22.5 10.4 10.6 38.3 17.2 28.0 14.5 18.9 23.4 35.6 18.8 23.2 20.3 35.8 11.3
Cats
09/24/2014Learning Object Detectors with Weak Supervision
45
And dogs
09/24/2014Learning Object Detectors with Weak Supervision
46
Summary
• Weakly supervised object detector learning
• Existing methods• Generative
• MI-SVM
• MI-CRF
• Future directions• Active learning (eg. OPTIMOL, NEIL)
• Current research: improving MI-SVM
• Open questions: part-based, multi-modal data, etc.
09/24/2014Learning Object Detectors with Weak Supervision
47