0.25
0.5
0.75
1.25
1.5
1.75
2
Large-scale and larger-scale image search
Published on Oct 09, 20126557 Views
The first part of this tutorial, dedicated to large-scale image retrieval, will first introduce the typical use-cases and the datasets used for evaluation of image search when considering an unsupervi
Related categories
Chapter list
Large / larger-scale image search Introduction00:00
General outline00:25
Image search01:11
Visual Search - 102:53
Visual Search - 203:19
Visual Search - 303:56
Scalability for the image search problem04:29
Datasets - Oxford5k/Paris6k08:31
Datasets - Holidays09:42
Univ. Kentucky object recognition benchmark10:44
Datasets - Stanford Mobile11:22
Large scale image search: Image description & matching12:16
General Outline12:31
Image description - 112:36
Image description - 213:30
Global descriptors14:33
Matching local descriptors [Lowe04]15:37
Local description: image detector16:27
Local image descriptor17:55
Geometric matching with local descriptors20:33
Large-scale image search: Bag-of-words and extensions21:37
General outline21:51
Direct matching: The complexity issue22:03
Bag-of-visual-words - 124:16
Bag-of-visual-words - 225:56
Video Google [Sivic & Zisserman’03]27:51
Inverted file29:04
Interest of the voting interpretation31:34
Inverted file – Complexity - 131:39
Inverted file – Complexity - 234:42
Inverted file – Boosting efficiency - 135:48
Inverted file – Boosting efficiency - 238:05
Large vocabularies: Assignment cost39:22
Large vocabularies with learned quantizer42:02
Bag-of-words: Another interpretation43:39
Compromise on vocabulary size: k=20000 - 146:34
Compromise on vocabulary size: k=200000 - 146:50
Impact of the vocabulary size on accuracy47:42
Compromise on vocabulary size: k=20000 - 250:18
But with a better matching method (HE) ...50:23
Compromise on vocabulary size: k=200000 - 250:25
Geometrical verification50:35
BOV search in 1M images - ranks53:09
Geometrical verification on a large scale - 154:49
Geometrical verification on a large scale - 255:11
Weak Geometry consistency57:12
Weak geometric consistency - 159:03
Weak geometric consistency - 201:00:12
Large scale: BOV search in 1M images01:01:30
Query expansion in visual search01:02:06
Bag-of-words: concluding comments01:06:10
Questions ?01:08:13
Towards larger scale - 101:08:36
Towards larger scale - 201:10:09
Larger-scale visual recognition: Novel aggregation mechanisms01:11:19
Outline01:11:22
Motivation for new aggregation mechanisms01:11:46
Motivation - 101:12:39
Motivation - 201:12:41
A first example: the VLAD - 101:12:42
A first example: the VLAD - 201:14:24
The Fisher vector: Score function01:14:50
The Fisher vector: Fisher information matrix01:16:37
The Fisher vector: Application to images01:17:47
The Fisher vector: Relationship with the BOV - 101:19:05
The Fisher vector: Relationship with the BOV - 201:19:11
The Fisher vector: Relationship with the BOV - 301:19:16
The Fisher vector: Relationship with the BOV - 401:20:32
The Fisher vector: Dimensionality reduction on local descriptors - 101:20:45
The Fisher vector: Dimensionality reduction on local descriptors - 201:21:38
The Fisher vector: Power-law01:22:37
Relationship between VLAD and Fisher01:24:46
Examples: Retrieval - 101:27:13
Examples: Retrieval - 201:27:15
Examples: Retrieval - 301:27:39
Examples: Retrieval - 401:28:08
Examples: Retrieval - 501:28:17
Packages for Fisher vectors01:28:25
Questions?01:28:48
Larger-scale visual recognition: Efficient matching01:32:32
General outline01:36:22
Efficient matching: Outline - 101:36:27
Finding neighbors01:36:34
The cost of (efficient) exact matching - 101:37:13
The cost of (efficient) exact matching - 201:38:18
Need for approximate nearest neighbors01:38:31
Efficient matching: Outline - 201:39:48
Locality Sensitive Hashing (LSH)01:39:55
LSH - partitioning technique01:41:24
What kind of hash functions/partitions?01:41:54
Hash functions - Structured vs Learned01:42:14
Multi-probe LSH01:44:00
FLANN01:45:15
Issue for large scale: Final verification - 101:46:35
Issue for large scale: Final verification - 201:47:25
LSH for binarization - 101:47:43
LSH for binarization - 201:48:22
LSH: The two modes - approximate guidelines - 101:48:43
LSH: The two modes - approximate guidelines - 201:50:10
Efficient matching: Outline - 301:50:32
Hamming Embedding01:50:41
ANN evaluation of Hamming Embedding01:51:32
Matching points - 20k word vocabulary01:51:33
Matching points - 200k word vocabulary01:51:36
Matching points - 20k word vocabulary + HE01:51:41
Efficient matching: Outline - 401:52:08
A typical source coding system01:52:15
Relationship between Reconstruction and Distance estimation01:53:24
Searching with quantization [J’11]01:54:04
Product Quantizer01:54:49
Asymmetric distance computation (ADC)01:56:21
Estimated distances versus true distances01:58:30
Combination with an inverted file system01:58:48
Performance evaluation01:59:43
Product Quantization: Some applications01:59:49
Concluding remarks02:00:41
Larger-scale visual recognition: Conclusion02:01:35
General outline02:01:38
Large-/larger-scale image search02:02:26
Large Scale Experiments - 102:03:25
Large Scale Experiments - 202:04:28
Very Large Scale Experiments02:04:29
Final note: Search vs classification02:04:32
General conclusions02:05:15