Deng 2009 - ImageNet

Routing Summary

This folder fully captures Deng, Dong, Socher, Li, Li and Fei-Fei (2009), ImageNet: A Large-Scale Hierarchical Image Database (CVPR 2009, pp. 248-255): the WordNet-based label space, the two-stage construction pipeline (search-engine candidates, AMT verification with adaptive per-synset consensus), the four audited properties (scale, hierarchy, accuracy, diversity), the comparison with TinyImage, ESP, LabelMe and Lotus Hill, three demonstration applications and the future-work plan. It contains 4 notes (paper is 8 pages).

Concept Map

ConceptNoteTypeDepends OnKey Result
SynsetDeng 2009 - ImageNet - OverviewdefinitionWordNet12 subtrees, 5,247 synsets, 3.2M images
Candidate poolDeng 2009 - Construction Pipeline and AMT Quality Controldefinitionsearch engines, WordNetover 10K candidates per synset
Confidence score tableDeng 2009 - Construction Pipeline and AMT Quality Controldefinitionat least 10 votes on a sampleharder synsets need more votes
Semantic difficulty from vote tableDeng 2009 - Construction Pipeline and AMT Quality Controltheoremconfidence tableno majority implies hard to illustrate
ImageNet hierarchyDeng 2009 - Dataset Properties and Comparison with Other DatasetsdefinitionWordNet IS-A147 dog categories
Average-image diversity measureDeng 2009 - Dataset Properties and Comparison with Other Datasetsdefinitionlossless JPEG sizesmaller size means more diverse
Property matrix (Table 1)Deng 2009 - Dataset Properties and Comparison with Other Datasetsconceptdatasets comparedonly ImageNet has all of 5 properties
NBNN distanceDeng 2009 - Applications and Future WorkdefinitionSIFT descriptorsfull resolution beats 32x32 NN-voting
Tree-max classifierDeng 2009 - Applications and Future Workdefinitionper-synset classifiersbeats independent classifiers at nearly every height

Notes

Sources

  • Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). ImageNet: A Large-Scale Hierarchical Image Database. CVPR 2009, 248-255. Deng 2009 - ImageNet

See Also