Explorable active learning method using Bayesian dual autoencoder for object detector and active learning device using same

摘要

It relates to an explorable active learning method using a Bayesian dual encoder for an object detector, (a) inputting test images to an object detector to generate cropped images, resizing the test images and cropped images, and resizing inputting the images to a data encoder to output data codes; (b) (b1) (i) input test images to object detector, apply Bayesian output embedding, resize activation entropy maps and cropped activation entropy maps, or (ii) input resized object images; apply Bayesian output embedding, (b2) input the resized activation entropy maps to the model encoder to output model codes; (c) (i) validate reference data codes, select specific test images with rare samples, update data codebook, (ii) validate reference model codes, and select specific test images with hard samples .