Today we can see a number of papers in which Deep Learning techniques are used to identify images that “are most likely” to improve an underlying model. In other words, Deep Learning is used to address Active Learning tasks.
One example is the following paper: https://arxiv.org/pdf/1904.00370.pdf
The strong point is that the approach works for classification, detection, segmentation and relevant tasks. The underlying idea is pretty simple: to have a model that for a given image can estimate how similar this image to the ones in the training set. Note, we don’t have to iterate over all the images in the training set to answer the questions.
Three following apps might be created to address the task.
I. Training app
An app takes as input an image project with training data and outputs a model that is capable of answering “how likely that a given image will improve the model?”
II. Inference app
An app takes as input
(1) an image project with unlabeled data
(2) a model that was produced by Training app
An App outputs a score for each unlabeled images, probably in a form of a tag
III. Image extraction app
An app takes as input
(1) an image project where images have a score associated. Scores are obtained as a result of running “Inference app”.
(2) target project name
App also allows
(1) to play with score threshold to understand the amount of images to be selected and to be added into training set
(2) visualize selected images in a “thumbnail view”.
App produces a new dataset within the “target project” which contains high score images (score is above a given threshold)
