I want to detect cars in aerial images from a low altitude(drones) with YOLACT. So far I've gathered data that has yet to be labeled.
The thing is, the original images vary in aspect ratios - 5472x3648 (3:2) and 3840x2160(16:9) respectively.
When resizing the images to the input size of YOLACT (550x550) the cars looked either squished or very elongated. I'm concerned that YOLACT won't be able to learn the masks of the cars properly as they look squished or elongated depending on their pose.
Therefore, my questions:
1.) Are aspect ratios besides 1:1 supported? Or even multiple aspect ratios like in my case?
2.) If not, should I just let them resize to 550x550 and hope for the best?
3.) OR, Should I crop them to 1:1 and lose information or use padding with black bars?
Besides that, I'm uncertain about cut off cars. Should I label them or only those who are fully present in the frame?
A lot of questions I know, but I would be very happy about your help
JunsukLee
commented
2 years ago
Hello everyone,
I want to detect cars in aerial images from a low altitude(drones) with YOLACT. So far I've gathered data that has yet to be labeled. The thing is, the original images vary in aspect ratios - 5472x3648 (3:2) and 3840x2160(16:9) respectively.
When resizing the images to the input size of YOLACT (550x550) the cars looked either squished or very elongated. I'm concerned that YOLACT won't be able to learn the masks of the cars properly as they look squished or elongated depending on their pose. Therefore, my questions: 1.) Are aspect ratios besides 1:1 supported? Or even multiple aspect ratios like in my case? 2.) If not, should I just let them resize to 550x550 and hope for the best? 3.) OR, Should I crop them to 1:1 and lose information or use padding with black bars?
Besides that, I'm uncertain about cut off cars. Should I label them or only those who are fully present in the frame?
A lot of questions I know, but I would be very happy about your help
Regards, Ruwen