 |
At its F8 developers conference Facebook not only revealed a number of new Instagram features, the company also talked about how it is using the billions of images on Instagram to train the world’s most accurate image recognition systems.
Training deep learning models for image and object recognition is typically a very labor-intensive task, as each training image has to be looked at and labeled by human workers. This is a serious limitation to the size of training image databases; however, Facebook has found a way to reduce human supervision in the training process by using images that are already labeled… with Instagram hashtags.
Its researchers used 3.5 billion Instagram images with approximately 17,000 hashtags to train deep learning models and the results have been encouraging.
A computer vision system that had been trained with one billion images and 1,500 hashtags, achieved 85.4 percent accuracy on the ImageNet benchmarking tool, outperforming the previous leading system by 2.3 percent.
It will be important to manage the disadvantages of less curated labels but the Facebook research shows that less supervised training of image recognition systems could be a step into the right direction, allowing for the use of much larger and databases and therefore improved image and object recognition and classification.
Translation: finding that photo you never tagged that’s buried miles deep in your archive might soon get a whole lot easier.
Articles: Digital Photography Review (dpreview.com)