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

Table 12

Training, updating, and retrieving time table of the Clothes CFOR system with the DeepFashion dataset.

Phase	Option	Running time
Offline: learning phase (note: models can be trained individually to increase training speed)	Training object detection model	∼6 training hours for detection model
	Training object classification models	∼20 training hours for all 4 models in the system: 1 region identification model and 3 category classification models for Top, Bottom, and Body regions, respectively
	Training attribute learning models	∼22 training hours for all 5 grouped attribute multitask classification models in the system, including shape, part, style, texture, and fabric
	Updating system	∼2.5 training hours per model for all mentioned models in the training phase (1 region identification model, 3 category classification models, and 5 attribute multitask classification models)
Online: retrieval	Identifying object regions, categories, and attributes	2 to 3 seconds per sample
	Retrieving	1 to 10 milliseconds per sample (may delay by the run-time system)