Hey, I want to use Dope with higher-resolution images, around 1000 × 1100. I’m wondering where I need to make changes in Dope. Here are some areas I’ve been thinking about:
Preprocessing:
Should I scale or crop my images to a square resolution before feeding them into training or inference? For example, 1000 × 1000.
Training:
During training, I noticed that the augmentation step crops a random 400 × 400 region from the image. Can I simply comment this out, or do I also need to adjust other parameters in the training code? My generated dataset already includes many images where the object is only partially visible, so additional cropping might not be necessary.
Inference:
Do I need to make any major changes here? I know that the belief maps are scaled to w/8 and h/8 — does this happen automatically?
Hey, I want to use Dope with higher-resolution images, around 1000 × 1100. I’m wondering where I need to make changes in Dope. Here are some areas I’ve been thinking about:
Preprocessing:
Should I scale or crop my images to a square resolution before feeding them into training or inference? For example, 1000 × 1000.
Training:
During training, I noticed that the augmentation step crops a random 400 × 400 region from the image. Can I simply comment this out, or do I also need to adjust other parameters in the training code? My generated dataset already includes many images where the object is only partially visible, so additional cropping might not be necessary.
Inference:
Do I need to make any major changes here? I know that the belief maps are scaled to w/8 and h/8 — does this happen automatically?