When I resize my dataset to 1333 800, I found that the small target detection effect is very bad, but after input with 4096 2048, cuda memory is not enough. Can you help me?
For inference, you could try to manually manage the memory or use a sliding window approximation. For training, I think that resolution is just too high.
When I resize my dataset to 1333 800, I found that the small target detection effect is very bad, but after input with 4096 2048, cuda memory is not enough. Can you help me?