try to be faster: use smaller network or reduce the input size?

[braindevices]

I have very large image 4000x3000, but the objects in it are quite small 20x40 to 150x300.
To get reasonable inference speed, I found that v5s+resize-to-1000 and v5n+resize-to-1500 gives me similar speed.

Now I want to retrain the model with my own data. I wonder usually which way is better?

1. use less parameter, but larger input image. Larger image keeps more detail to let the network to work on.
2. use more parameter, but smaller image. Although there are more parameters, but the input has worse quality.

[braindevices]

to me, intuitively, I think the 1st is better. Because network may learn a smart way to ditch redundant information. But then the question will be how easy to reduce the network size further? sparsify it?