Consistency-based Semi-supervised learning for Object ...Type of object detector: quick review...
Transcript of Consistency-based Semi-supervised learning for Object ...Type of object detector: quick review...
CAP 6412 Advanced Computer Vision, Spring 2020
Consistency-based Semi-supervised learning for Object Detection
Authors: Jisoo Jeong, Seungeui Lee, Jeesoo Kim, Nojun Kwak
Venue: Advances in Neural Information Processing Systems 32 (NIPS 2019)
Presenter: Ishan Dave
1
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Outline of this presentation
● Understanding the problem and terminologies● Existing Approaches to solve the problem● Inspiration of the proposed work● Proposed method● Technical Contribution● Results● Conclusion
2
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Understanding the problem and terminologies: Object Detector
Image Courtesy: https://towardsdatascience.com/evolution-of-object-detection-and-localization-algorithms-e241021d8bad
3
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Type of object detector: quick reviewSingle stage Obj Detector
Two stage Obj Detector
Region Proposal Network Classifier
Eg. SSD, YOLO
Eg. Faster RCNN
4
Liu, Wei, et al. "Ssd: Single shot multibox detector." European conference on computer vision. Springer, Cham, 2016.Redmon, Joseph, and Ali Farhadi. "Yolov3: An incremental improvement." arXiv preprint arXiv:1804.02767 (2018).Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE international conference on computer vision. 2015.
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Understanding the problem and terminologies: Motivation
Supervised LearningObject
DetectorModel
Classification and Localization Loss
Semi-Supervised Learning
Object DetectorModel
Classification and Localization Loss for
labeled data
But what losses for Unlabled data?
5
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Any questions till this point?
✓ Understanding the problem and terminologies✓ Object Detector✓ Different Level of Supervision✓ Motivation
❏ Existing Approaches to solve the problem
6
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
A common approach in Semi-supervised learning
Object DetectorModel
Classification and Localization Loss
TRAINEDObject
DetectorModel
Labeled data
Prediction“Horse” at [245, 240, 50, 50] with 0.93 confidence
Confidence threshold
Inference Mode
Training Mode
“Pseudo Labels”
“Self-Training approach”
7
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Consistency regularization
Classifier Model
Original (x)
Perturbed (x’)
Loss
f(x)
f(x’)
State-of-the-Art in Semi-Supervised Classification!
8
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Consistency regularization
Localization Model
Original (x)
Perturbed (x’)
f(x)
f(x’)
Hard to find one to one correspondence!
9
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Image courtesy https://www.petmoo.com/dogs/golden-retriever-puppies/
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Proposed Work “CSD”
1010
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
CSD-Single-Stage Detector
11
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
CSD-Two-Stage Detector
12
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Loss function
1313
Overall Loss
Supervised Loss
Classification Localization
Consistency Loss
Classification Consistency
Localization Consistency
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Consistency Loss
For a pair of Bounding Box
Average classification loss for all Bounding box pair
Prediction of Localization:
≆
Average classification loss for all Bounding box pair
Total Consistency Loss
Classification Localization
14
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Overall loss for Object Detector
Weight Scheduling for “Ramp-up and Ramp-down Strategies”
15
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Background EliminationConsistency loss computed with all candidates will be easily dominated by background objects
Image courtesy: https://www.learnopencv.com/wp-content/uploads/2017/10/object-recognition-false-positives-true-positives.jpg
16
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Example:For background object (in blue) the detector predicts some prediction value: “Background” as “Dog”: 0.3, 0.2, 0.3, 0.1, 0.2----> should be ‘0’
Loss: 0.3+0.2+0.3+0.1+0.2= 1.1
“Dog” as “Dog”: 0.9, 0.8---->Should be ‘1’Loss:0.1+0.2= 0.3
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
ExperimentsDatasets:
1. PASCAL VOC: 20 Classesa. VOC07- 5k imagesb. VOC12- 11.5k images
2. MSCOCO: 80 Classes
17
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Experiments
18
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Metric:
Mean Average Precision (mAP): AUC of Precision-Recall curve and take average across classes
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Results
1. Using CSD method improves the mAP
2. Combined classification and localization consistency provides better mAP than the individual
3. Combined BE, classification and localization consistency provides the best mAP
4. Improvement in 2-stage object detector is lesser than that of 1-stage object detector
19
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Results-Effect of Unlabeled data with different distribution
2020
20 Classes
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Results: Consistency loss without unlabeled data
2121
Understanding the Problem
Existing methods
Inspiration of the work
Proposed Method
Technical Contribution Results
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
Limitations 1. Not significant improvement on 2-stage Object Detector (~0.8%)
2222
2. Unlabeled data should have a similar distribution to the labeled data
Consistency-based Semi-supervised Learning for Object Detection (NIPS-2019)CAP 6412 Advanced Computer Vision, Spring 2020
ConclusionUsing Semi-supervised technique- Consistency regularization this paper improves the performance of the object detector using unlabeled data
2323