YOLO on a custom (transfer tranied) ImageNet model

[deadbeef101010]

YOLO on a custom (transfer tranied) ImageNet model

Hello we (son and self) are working on a high-school project to detect mites and then locate those mites on diseased bees. Of course we are assuming many things are ideal to keep things simple. We are far from experts and are seeking some pointers.

We plan to do transfer learning (TF) on ImageNet and KNN classification of the mites, then wanted to use YOLO to locate those mites in an input image. However we have questions as follows:

• We tried TF as a test on several mite images. However, wanted to know if all the image sizes we feed into the classifier need to be the same size?
• Once trained, and in order to classify do we need a similar input test image size?
• Does the test image need to be primarily composed of the object being classified?

The next step would then be to use the trained ImageNet model on the new classification to locate the mites in an input image that could have other objects too, however, the YOLO API does not allow to specify our new classifier. Is this even possible? Are we on the correct path?

Thanks for your help, sorry if this is not the place for this---if not, please indicate where to post instead.

[Swizz]

I am also interested in using a custom YOLO pre-trained model to locate only specifics objects on an image.

Any update on this ?

[joeyklee]

Hi @Poohgit!

The short answer:

• ml5 is likely not the best tool to achieve this level of complexity.

The longer answer: +It seems you can do transfer learning on YOLO but this requires some more python tooling and probably a GPU? https://github.com/ultralytics/yolov3/wiki/Example:-Transfer-Learning

[joeyklee]

Just making a note that for custom trained object detectors I believe a functionality that one can do in RunwayML. Maybe worth looking into there. Closing this issue now. Thanks!