Download presentation
1
Predicting Matchability - CVPR 2014 Paper -
Min-Gyu Park Computer Vision Lab. School of Information and Communications GIST
2
Intro (1/3) Point feature extraction and matching
Initial step for various computer vision algorithms SFM, object recognition, feature-based tracking… SIGGRAPH 2014, Feature Matching with Bounded Distortion
3
Intro (2/3) Millions of images!!!
Feature extraction and matching step can be the bottleneck for large-scale applications SFM from collected images from the web, Bundler, Microsoft Millions of images!!!
4
Intro (3/3) Possible approaches to reduce the time complexity
Reduce the number of images Make a cluster of images and select a representative image Developing fast matching methods Approximate Nearest Neighbor (ANN), k-NN, etc Extract fewer numbers of features This paper belongs to this category
5
Predicting Matchability (1/3)
Predict matchable features priori to matching! To remove un-matchable features The simplest approach might be, Reject if the detector response is weak Cornerness is less than a user-defined threshold value However, repeatable and well-localized points do not guarantee they are “matched correctly”
6
Predicting Matchability (2/3)
Goal Train a classifier to predict matchable features As well as to rule out un-matchable features! Training step Classifier is trained in the Random Forest framework SIFT descriptors (extracted from training images) Positive samples (485,000) Negative samples (485,000)
7
Predicting Matchability (3/3)
Testing step Run down each tree in the random forest Proposed results
8
Experiment Detection performance
Accuracy of prediction in terms of ROC curves Qualitative evaluation SFM results with predictable matches
9
Detection performance
ROC curves for three datasets Green: proposed Blue: DoG threshold Red: random selection (d) Confusion bars
10
SFM with Proposed Method
The proposed method better recovers the shape of the object properly
11
Q & A
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.