Training custom model for scene classification?

[gillonba]

Hello, This is an excellent project, and I am impressed how easy it is to use, and how well it works! Now I am wondering about some things I can do with it though; I have built a custom model for object detection within an image, but am wondering if I can do the same for scenes, in order to do more specific classification of detected objects.

For example, if I trained a model to pick out cars in an image (or use an existing one), could I then grab the cars from the image and feed them into a scene classifier trained to identify the specific make/model of car? I am worried that if I train an object detector with each make/model it will then return many results on each car in the scene, leaving me to do an awful lot of sorting. Is there a better way to achieve this goal?

Great work either way!

[robmarkcole]

This is currently a feature request. There are 2 parts to your request:

• deploying custom classifiers
• chaining models, i.e. detect car, then classify it

[gillonba]

Essentially, yes. Is there any meaningful difference between the scene classifier and an object detector that only returns the most prominent object?

[robmarkcole]

I expect you can get better % accuracy from the classifier. With obj detection there are all these issues around small object detection which makes them less accurate. Also training requirements are simpler for classifier, and you dont need to annotate the objects

[gillonba]

Are there any updates on this? Is a custom classifier likely to happen? I can handle the chaining myself if need be. This is a great project but unfortunately it is looking pretty quiet lately