Large-Scale Coarse-to-Fine Object Retrieval Ontology and Deep Local Multitask Learning
Table 1
Contributions of CFOR in the offline phase and its comparison with DeepFashion [8].
Criteria
Object ontology (categories/attributes)
Deep learning model
Imbalanced data problem solver
Updating system
Searching method
CFOR system
Object ontology can be implemented on arbitrary objects with flexible modifications. It is not just for fashion. Ontology is a semantic hierarchical tree. It consisted of three levels: (i) Region level (ii) Category level (iii) Attribute level Region level consisted of some parts of objects; each part is linked to the category level. Category level consisted of some fine-grained groups; each member of the group is linked to the attribute level. Attribute level consisted of two coarse-grained groups: visual concept (color, shape, and texture) and specific attribute concept (fabric, part, and style). It is easily implemented on another kind of object with flexible modifications. It is not just for fashion.
ResNet-101 is used in category classification NASNet v3 is used in attribute classification
Based on Matthews’ correlation coefficient
Based on transfer learning. Trains smaller models based on ontology and updates the system in a flexible manner in each model
The authors do not use the terminology “ontology.” Dataset is organized based on a hierarchical tree. It is implemented just for fashion. It consisted of two-level trees: (i) The first level consisted of category groups (50 categories) (ii) The second level consisted of 5 attribute groups (texture, fabric, shape, part, and style). It does not have color attribute (It treats concepts and attributes as independent tasks of classification).