Annotation With Only YES Or NO! A Very Efficient One-bit Annotation Proposal!
3 main points.
✔️ We propose a method for combining semi-supervised learning and one-bit annotation
✔️ Substantial accuracy improvement in image classification tasks at the same cost as semi-supervised learning
✔️ Flexible method that can be applied to object detection and segmentation
One-bit Supervision for Image Classification
written by Hengtong Hu, Lingxi Xie, Zewei Du, Richang Hong, Qi Tian
Cite as: arXiv:2009.06168 [cs.CV](or arXiv:2009.06168v2 [cs.CV] for this version)
Subjects: Computer Vision and Pattern Recognition (cs.CV)
First of all
Over the years, deep learning technology is starting to be applied in various fields. The reason why deep learning techniques are used so much is being used to such an extent mainly because of its vastly better accuracy compared to classical methods and its flexibility.
As Corona explodes around the world, traditional business models will no longer be viable, and we will see a shift to a more digital world. In this digital shift, the importance of deep learning technologies will continue to grow.
However, deep learning techniques suffer from several major problems. One of them is the need for a large amount of data for training and the high annotation cost associated with it. The cost of collecting large amounts of data is becoming cheaper with the advent and proliferation of a wide variety of sensors. However, labeling the collected data remains expensive, as it is proportional to the number of data. Especially for datasets with a large number of classes, such as Imagenet, the annotation cost can be very high.
In the paper presented here, we propose a new low-cost annotation method called one-bit annotation. By combining this annotation method with semi-supervised learning, we achieve high accuracy in the image classification task.
To read more,
Please register with AI-SCHOLAR.OR
Categories related to this article