However, I am trying to use this dataset to hopefully use it on more complex ones, such as the second image. Because the plant village dataset is already set to draw bounded boxes, since it is the entire image, I was hoping to use it as inputs for yolo. However, when I see examples of yolo, bounding boxes for objects such as a human are a small part of the image, which also has complicated backgrounds. Is it advisable to pass in these plant village images as having "bounded boxes", which is basically the entire image? Does this require a different algorithm or technique?
- In general, if you want to transfer knowledge learned on one dataset to another, the more similar they are, the more likely it will work. If the datasets are very dissimilar, this may be very hard and they may not be any simple solution.
- There is no point in having entire pictures surrounded in bounding boxes. In such a scenario, your algorithm would have nothing to learn, it would just need to always return an image-sized box as a result. It is like learning your algorithm to always return "true".
- For possible solutions, you should either re-define your problem or look for better data. Maybe instead of classifying a diseased plant, train an algorithm to find the parts of the plant that are infected (this would appear in both domains), in such case you could decide that the plant is diseased if it has diseased parts.
- What you need is the real-life labeled data (images like the one below). Even if you can learn something from the "clear" pictures and transfer it to the images below, it will be a gamble and it does not give you much guarantees on the accuracy in the wild. At worst, you need a big and representative test dataset consisting of real-life, labelled images to validate your results.