Image resolution don't improve detection

[cespedesk]

So, I was ruining inference in some objects and I notice that even when the camera is high resolution it does not detect far objects. then I created a small region of interest and ran the inference on it. Then it detected objects from very far. It is like it cares of how much of the image is being occupy by the object. Is there any way to detect far objects taking advantage of FHD or even UHD cameras?

[AastaLLL]

Hi,

This depends on the model you used.

For example, in YOLOv3, the supported input image is 608 or 416.
So even you feed a high-resolution image, the framework will rescale it into the model input size for prediction.

Thanks.

[cespedesk]

So, what would be the framework with the higher resolution input? I what to detect objects from very far.
I also learned that the input resolution can be changed, if so how cam I change the input resolution in the costume dataset training of Hello AI Wold?

[AastaLLL]

Hi,

This depends on the model you used.
You can find a model that supports the resolution for your use case.

If you want to train one on your own, could you share which model do you want?
So we can suggest the corresponding modification for you.

Thanks.

[cespedesk]

I am using the ssd-mobilenet from jetson-inference/pytorch-ssd.md at master · chiou17/jetson-inference · GitHub. Is this what you are asking me?

[cespedesk]

I found the answer in another post, thanks for your attention.