I’m working with:
- Jetson Xavier NX
- Deepstream 5.0
- Python and RSTP cameras
I’m using the model in samples/models/PrimaryDetector/resnet10 as it is to identify cars that enter a forbidden area.
I have seen that the model works very well on cars that are very close to the camera (like in the sample videos: around 3-4mt) or kinda far away as I tested with my RSTP camera (50mt). However, it tends to miss detections around some distance in the middle (10mt or so).
Is there a recommended distance in which this pretrained model (As it is, with no retraining) works best?
In our training data, we don’t collect it for a particular distance.
So it should work for the various cases.
But it might be good or bad in some distance due to the DNN learning.
In general, we recommends user to set bbox > 4.
Based on the information, you can roughly find the expected distance with the width/height of a real car.
Ok, I understand.
But now, what do you mead by: set bbox > 4?
Sorry for the non clear statement.
We recommends that a valid bbox should have width/height more than 4 pixels.
You can use this information to calculate the ideal distance.