Graph-Based Discriminative Learning for Location Recognition
read more
Citations
NetVLAD: CNN Architecture for Weakly Supervised Place Recognition
Benchmarking 6DOF Outdoor Visual Localization in Changing Conditions
NetVLAD: CNN Architecture for Weakly Supervised Place Recognition
Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization
Image-Based Localization Using LSTMs for Structured Feature Correlation
References
LIBLINEAR: A Library for Large Linear Classification
Video Google: a text retrieval approach to object matching in videos
Object retrieval with large vocabularies and fast spatial matching
Three things everyone should know to improve object retrieval
Related Papers (5)
Frequently Asked Questions (12)
Q2. What future works have the authors mentioned in the paper "Graph-based discriminative learning for location recognition" ?
Finding a way to automatically adjust learning parameters or synthesize the results from different clusters is an important issue, and an interesting direction of future work.
Q3. What is the main contribution of the approach?
A main contribution of their approach is to combine the power of discriminative learning methods with the rich structural information in an image graph, in order to learn a better database representation and to better analyze results at query time.
Q4. What is the way to use the simple fall back strategy?
In their experiments, the authors use the simple fall back strategy by default, and separately evaluate a combination of averaging and interleaving as a stronger form of tf-idf regularization.
Q5. What is the way to regularize the ranking of query images?
as a simple strategy, for query images where all models give a probability score below a minimum threshold Pmin (0.1 in their tests), the authors fall back to tf-idf scores, as the authors found low probability scores unreliable for ranking.
Q6. what is the conditional probability of a picture matching?
The last line in the derivation above relates P ′b to Pb via an update factor, (1− PbaPb Pa)/(1−Pa), that depends on Pa (the probability that the top ranked image matches) and Pba (a conditional probability).
Q7. what is the conditional probability that image b matches the query?
Pb − P (Xb = 1|Xa = 1)P (Xa = 1)1− Pa= Pb − PbaPa1− Pa = Pb ( 1− PbaPb Pa 1− Pa ) (1)where Pba = P (Xb = 1|Xa = 1) denotes the conditional probability that image b matches the query given that image a matches.
Q8. What is the way to improve the quality of the matching?
These matches are sufficient for their method (though to improve the quality of the matching, the authors can also run structure from motion to obtain a point cloud and a refined set of image correspondences).
Q9. what is the probability of image b matching the query?
The update factor in Eq. (1) has an intuitive interpretation: if image b is very similar to image a according to the graph (i.e., Pba is large), then its probability score is downweighted (because if a is an incorrect match, then b is also likely incorrect).
Q10. Why do the authors believe that the global ranking is contaminated by the low-degree nodes?
The authors believe this is due to the nature of image graphs for unstructured collections, where some nodes have many neighbors, and others (e.g. very zoomed-in images) have only a few; training and calibration for these low-degree nodes may result in models that overfit the data and contaminate the global ranking.
Q11. What is the way to determine the pose of the query image?
If the authors have a 3D structure from motion model, the authors can then associate 3D points with matchesin the query image, and determine its pose [18].
Q12. How do the authors use this information to verify the location of a query image?
The authors make use of this structural information in a bag-ofwords-based location recognition framework, in which the authors take a query image, retrieve similar images in the database, and perform detailed matching to verify each retrieved image1until a match is found.