Revision history [back]

Let me reply to your questions, since I got quite the background using the specific technique

1. Part one of question one is complete bullshit. Your ROI of your objects should correspond to the physical dimensions in real life and thus that width/height ratio, but not correspond to the image dimensions in which it was captured.
2. Second part of question one is partially right. In the createsamples step, it will basically take your annotations and rescale them to the model dimensions. If your ratio is quite fluxuating, then all your samples will be rescaled to a fixed size ratio, introducing undesired deformations. Think of deforming a flat rectangle to a square model size. It will push together all data in some direction.
3. Like said, answer on question 2 is that ratio should be about same during annotation. Enforcing the ratio to all samples will be done by the opencv_createsamples tool.
4. No - negative images can be as random as you want!

Let me reply to your questions, since I got quite the background using the specific technique

1. Part one of question one is complete bullshit. Your ROI of your objects should correspond to the physical dimensions in real life and thus that width/height ratio, but not correspond to the image dimensions in which it was captured.
2. Second part of question one is partially right. In the createsamples step, it will basically take your annotations and rescale them to the model dimensions. If your ratio is quite fluxuating, then all your samples will be rescaled to a fixed size ratio, introducing undesired deformations. Think of deforming a flat rectangle to a square model size. It will push together all data in some direction.
3. Like said, answer on question 2 is that ratio should be about same during annotation. Enforcing the ratio to all samples will be done by the opencv_createsamples tool.
4. No - negative images can be as random as you want!