Thanks for the tutorials I've been looking to get my head around OpenCV and your videos have helped alot. I was wondering how transferable the haar cascade trained on your watch is? For example If I had 2 watches/ objects that were slightly different due to colouring or textures would the 1 haar cascade work for both or can I use the cascade.xml trained from object 1 to speed up the training of detecting object 2?
Maybe my question is simply how different would a watch need to be from the trained image to not get detected and can you build on existing haar cascades to improve their detection of those variants ?
Michael
You must be logged in to post. Please login or register an account.
In our example, when we did create_samples, we used the single image to create all of those examples. If you wanted to detect general analog watches, rather than using create_samples on just that image, you'd use it on thousands of variations of watches (getting maybe from imagenet). I doubt the Haar Cascade used in the series will detect watches that are not extremely similar to mine since that was the only watched trained on.
-Harrison 8 years ago
You must be logged in to post. Please login or register an account.
So a middle ground might be to get a few more examples of positives and run create_samples instead of trying to extract and label your dataset of positives?
Would the choice of negatives matter much? As in the context, if I wanted to detect a surfboard for example, it would (to me) make more sense to have negatives include things like beach, sea, sand or is that not the worth the hassle ?
-mobcdi 8 years ago
You must be logged in to post. Please login or register an account.
You could just keep the negatives we used or find some others, negatives just need to not have whatever you're tracking, that's all. You could try various watches and styles, maybe something like 5 or so total that seem to have traits that are seen in the rest of watches and see what happens. Ideally though, if you want a general analog watch detector, you will want to have hundreds of examples, if not thousands. At least in the lower numbers, you can relatively easily test to see what the minimum is you can get by with via trial and error and simply training overnight.
I would wager, for the surfboard example, it wouldn't make much if any difference, and likely result in a mistake where you have some negatives that might actually have surfboards.
-Harrison 8 years ago
You must be logged in to post. Please login or register an account.
As for creating the positive images, if it were possible to track the object using other methods could individual frames with the correct roi count as the positive images or would there be a need to have more variety in the positive images like different lighting conditions, rotations etc ?
-mobcdi 8 years ago
You must be logged in to post. Please login or register an account.