1
$\begingroup$

I am training an algorithm to identify weeds within crops using the YOLOv5 algorithm. This algorithm will be used in the future to identify weeds in images collected by unmanned aircraft (drones) after making an orthomosaic images. Using the open-source LabelImg software, I am labeling images for object detection that were collected with both UAV and hand-held digital cameras. Using both platforms, I collected many images of weeds that will need to be identified.

My question is this: Does it make sense to collect training samples from the hand-held digital camera, since it will be of much higher resolution than the UAV imagery (and thus not used for future imagery collections after the model is trained)? My initial thought is that it would be best to only use the UAV imagery, since it will be the most similar to what will be collected in the future. However, I do not want to throw out the hand-held digital imagery if it could help in the image classification process.

$\endgroup$
1
$\begingroup$

I think this can only be used for pretraining/some kind of transfer learning. This would be useful if the ratio of real training data to pretraining data is really low. You could then pretrain on digital data, and fine-tune on your UAV data.

How much of this is really useful I can't really say, this depends on how close the digital data is to UAV data. If it is significantly different, you are training on a different distribution and sample space than your UAV images, which is pointless.

| improve this answer | |
$\endgroup$
  • $\begingroup$ That was my thought. It would be most useful to train on data that would be the most similar to the output of the end goal, so I chose to use the UAV data. Thanks. $\endgroup$ – ihb Sep 18 at 14:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.