6. What’s the problem?
• Semantic difficulties: fine-grained differences
Image credits: http://ai.stanford.edu/~jkrause/cars/car_dataset.html
NGUYEN ANH TUAN 東京大学・情報理
工・修士2年生
7. But for search problem?
Image credits: http://ai.stanford.edu/~jkrause/cars/car_dataset.html
Query Database
NGUYEN ANH TUAN 東京大学・情報理
工・修士2年生
8. But for search problem?
Image credits: http://ai.stanford.edu/~jkrause/cars/car_dataset.html
Query Database
0.1
0.5
0.2Ranking problem
with a variation of
fine-grained
changes
NGUYEN ANH TUAN 東京大学・情報理
工・修士2年生
9. But for search problem?
Image credits: http://ai.stanford.edu/~jkrause/cars/car_dataset.html
Query Database
0.1
0.5
0.2Find visual representations
to capture all fine-grained
local information in images
NGUYEN ANH TUAN 東京大学・情報理
工・修士2年生
14. Image matching = Feature matching
• Feature matching→Nearest Neighbor Search
– Inverse Search with Inverted Indices
– Compressed data for better memory usage [3]
Feature
extraction
Feature
aggregation
Feature
matching Re-ranking
Preliminary
results
Final
results
[3] H. Jégou, M. Douze, C. Schmid, Product
quantization for nearest neighbor search., IEEE
Trans. Pattern Anal. Mach. Intell. 33 (2011) 117–
28.Data CompressionNGUYEN ANH TUAN 東京大学・情報理
工・修士2年生
15. Verification
• Geometry verification
– RANSAC methods [4]
– Reduce the number of good inliers
Image credits: http://ai.stanford.edu/~jkrause/cars/car_dataset.html
Feature
extraction
Feature
aggregation
Feature
matching Re-ranking
Preliminary
results
Final
results
[4] M.A. Fischler, R.C. Bolles, Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,
Commun. ACM. 24 (1981) 381–395. NGUYEN ANH TUAN 東京大学・情報理
工・修士2年生