View-Based Maps

EC EN 631

Robotic Vision

Matching Views

• Compact Randomized Tree Signatures

• Prefilter for Place Recognition

• Geometric Consistency Check

• Experiments Results

Compact Randomized Tree Signatures

• Its a feature descriptor like SIFT, SURF or ORB
• Small dataset is chosen
• Base keypoints are extracted with a fixed pattern or FAST with some constraints
• Random Tree classifier is trained

Compact Randomized Tree Signatures

Text

Text

View-based map

Compact Randomized Tree Signatures

• Signatures are sparse and take lots of computational time
• Use PCA or random orthogonal vector multiplication to make it dense
• This feature representation is very fast than SIFT
• Thus for a new key point not in the base set, a new signature can be created

Vocabulary Trees

• Similar to the bags-of-word model
• Used for large vocabulary

• Tree construction

http://www.cs.utexas.edu/~grauman/courses/fall2009/slides/lecture16_bow.pdf

Vocabulary Trees

Tree: Training

http://www.cs.utexas.edu/~grauman/courses/fall2009/slides/lecture16_bow.pdf

Vocabulary Trees

Tree: Recognition

http://www.cs.utexas.edu/~grauman/courses/fall2009/slides/lecture16_bow.pdf

Geometric Consistency Check

1. Features matching (left image to left image)
2. Pick 3 candidates and generate relative motion hypothesis
3. Project the 3D points from one view to other and count inliers
4. Keep the hypothesis with best inliers
5. Do nonlinear estimation of the relative pose

• Using the binomial distribution for inliers a rejection filter rejects the false positives

Bigger Picture

Experiments

Experiments

Experiments

Thank You!