Blog

Deep learning basing on EM

Published: November 20, 2018

learning settings Training settings for object detection using CNN models. (a) is supervisedlearning which uses instance-level labels (red bounding boxes in the image). (b) isweakly-supervised learning where only image-level labels (”train”) are used. (c) is mix-supervised learning where both image-level labels (”train”) and instance-level labels(red bounding boxes) are used. (d) is semi-supervised learning where instance-levellabeled images and unlabeled images are used

Expectation Maximization algorithm

Published: May 19, 2018

I. Expectation Maximization

How to reduce the prediction time of deep learning models

Published: March 13, 2018

When we need speedup the prediction time?

Quadrilateral detection from polygon or segmentation masks

Published: November 13, 2017

In some applications we need to detect the quadrilateral objects in the image, for example detecting a document, an ID card, or anything whose the shape is quadrilateral. By using some techniques such as contour detection, edge detection, connected components etc. we can identifiy the object but it is not always a 4-edges polygon.