Hi all, does anyone have experience in using ImageNet? It is an image database that is organized according to the WordNet hierarchy and has over 14M images. I am wondering if anyone has used this database to train image recognition models / scrapers and how helpful it's been? Thanks!
9 November 2019 10:47pm
Hi, deep learning frameworks like Pytorch and Tensorflow come with state-of-the-art image recognition models (VGG, GoogleLeNet, ResNet, Inception etc.) already pre-trained on ImageNet, so one can just download and use them straight away.
This provides a solid starting point as these models have already learnt how to classify objects really well. Transfer learning can then be used to fine-tune them for a specific task like identifying ivory in photos. This just requires that you tweak the existing model a bit and train it on a smaller custom dataset of the image’s/categories you would like to classify.
Hope that helps!
9 January 2020 7:44pm
Really helpful, @adnortje ! And thanks for listing out some of the latest and greatest image recognition models. Do you know of any programs that are using GoogleLeNet for wildlife image recognition, by chance?
13 January 2020 4:31am
Google’s Wildlife Insights, mentioned in another thread, can classify 614 different animal species.
It uses the Inception-V4 model as a basis, which is similar to GoogLeNet, and tunes it on camera trap images (https://www.wildlifeinsights.org/about-wildlife-insights-ai).
André Nortje
Stellenbosch University