Large-Scale Coarse-to-Fine Object Retrieval Ontology and Deep Local Multitask Learning

Table 2

Contributions of CFOR in the online phase and its comparison with DeepFashion [8].

Criteria	Query	Retrieval process	Indexing method	Retrieval results
CFOR system	Image + optional semantic information (categories and attributes) extracted automatically from an image	Conducted by deep networks based on object ontology at three levels: region, category, and attribute levels	Quantized inverted indexing is operated by object ontology	Object ontology supports achieving retrieval results.; Retrieval results are based on deep global features and attribute vector.; Query expansion is used to improve the performance of the retrieval system.
FashionNet [8]	Image	Conducted by deep networks and landmark points (also built up by deep networks)	Inverted indexing	Result obtained is similarity retrieval only.