250 likes | 377 Views
Mid-Presentation Wei Li ENGN 2560 April 16,2013. Topic & Paper. Paper: Learning to Match Images in Large-Scale Collections. (Song Cao and Noah Snavely ) ECCV Workshop on Web-Scale Vision and Social Media , 2012 . http ://www.cs.cornell.edu/projects/matchlearn/. Problem Recall.
E N D
Topic & Paper Paper: Learning to Match Images in Large-Scale Collections. (Song Cao and Noah Snavely)ECCV Workshop on Web-Scale Vision and Social Media, 2012. http://www.cs.cornell.edu/projects/matchlearn/
Problem Recall Introduction: Large Scale Classification Problem Discover the visual connectivity structure from large image datasets. Goal & Expected Results: Determine Match Pairs & Non-Match Pairs. From an unknown large dataset. Using a small training dataset.
Steps & Outcome • Representation • of images • High similar pairs • Train the model • Testing Coding Testing • Step 1: Build Tf-IdfBoW model via SIFT feature for training dataset • Step 2: Find high similarity pairs for training SVM weights • Step 3: Using L2-regularized L2-loss SVMs to train the pair. • Step 4: Testing
Outcome 1: Build BoW model for training dataset <0.2 500/700 SIFT features (L2-Normlization) K-means cluster Tf-Idf weight
Outcome 1: Build BoW model for training dataset SIFT features (L2-Normlization) K-means cluster Tf-Idf weight
Tf-IdfBoWmodel 500 different features 10 pictures
Outcome 2 Get the similar pairs Sum of the difference High Similar if d > 1.5~1.8 median (d) Low Similar if d < 0.7~0.5 median (d) Abandoned if 0.7 median(d)< d < 1.5median(d) Use the distance of two images represented by Tf-IdfBoW Model. Set a proper threshold.
Match Image A & B
Non-Match Image A & C
Outcome 3 Train the model C constant: 0.3 0.5 0.8 1.0 Get the weight
Outcome 3 Train the model Get the weight
Outcome 4: Testing Notre Dame Training Dataset: 50 pairs/100 pairs 1/2 Match Pairs vs 1/2 Non-Match Pair 1/3 Match Pairs vs 2/3 Non-Match Pair Testing Dataset: 5000 pairs 1/2 Match Pairs vs 1/2 Non-Match Pair
Outcome 4: Testing General Result _ _ + + + + _ _ Precision = 61.3% Accuracy = 61.3% 2 _ _ + + + + _ _ Precision = 58.0% Accuracy = 59.0%
Outcome 4: Testing General Result _ + + _ Paper: TPR TNR : 0.4 ~ 0.8 _ + + _
Outcome 4: Testing Cluster Size Constant Parameter
Outcome 4: Testing Cluster Size 500 700
Outcome 4: Testing Constant Parameter TP FN FP TN
Other Observation Outcome 4: Testing It classifies the Match pairs better than Non-Match pairs. For k-means cluster, the initial center matters a lot For a small training set, proportion of Match Pairs and Non-Match Pairs is important. The training set affects the result a lot.
Outcome 4: Testing Match Pairs Non-Match Pairs Difficult!
Future Improvement Coding Other methods. ….. • Instance • Category Testing SIFT features (L2-Normlization) K-means cluster Tf-Idf weight Distance Threshold SVM (Hyperplane) ……
Conclusion 1. Goal is achieved: It can successfully classify the dataset via a small training dataset for an unknown dataset 2. Accuracy Estimation: High similarity pairs Numbers of visual words Weakness of SVM K-means cluster (sensitive to initial centers) ……
Time Schedule Week 6 (Apri 15-21): • BoW Model or something else Week 7 (Apri 22-28): • Do something if there is something which can be improved Week 8 (Apri 29-May 5): • Do something if there is something which can be improved Week 9 (May 6-12): • Reports &Summary May 14: • Final-Presentation