The device must be chosen correctly, it should be compatible with the supported devices listed in the installation guide.
Considering the capacity of your device, you should run a certain number of models and cameras. The number of models that can be run on devices depends on the size of the model that will be used for inference, the memory of the device, camera resolutions, and hardware capabilities.
For most cases, the following devices will support:
Jetsons, up to 3-4 cameras and 3-4 models should be run.
T4 GPUs, 10-12 cameras, and 15-16 models should be run.
While uploading the models on the devices, make sure that there is enough GPU space available. Uploading models one at a time will make sure to utilize the memory as much as possible.
The minimum camera resolution should be over 1080p
1080p is an ideal resolution for inference speed and accuracy.
The camera positioning is crucial for good model performance. The view frames obtained from the camera should be similar to the images used for annotation and training of the model to prevent domain gaps. Object positioning, scale, lightning of the environment, and background should be selected similar to the training set. If training data on hand is very different from the camera view, please use synthetic data and augmentation tools to generate images that can generalize the data on hand to production images.
See here to get a Device Overview