I am currently running a detection and tracking program on my Jetson Nano 2GB. The detection is made with the proposed ssd-inception-v2, which works great. The tracking is made with a Kalman Filter and a ByteTrack implémentation.
My question is about memory usage. Idle, apparoximately 700MB are used. When I run my program, this number goes up to 1.8 GB, that is, almost full capacity.
I would like to add another element to my program (that is, tracking by Re-id), but I am afraid the memory will saturate.
Can someone explain me why such a discrepancy in memory usage between “idle” and “in use” situations ?
Is there any way to limit this, so that I can enrich my program ?
I have found about this option for Tensorflow : (gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.4). But, if I am not mistaken, TRT is just not Tensorflow, although it is compatible with it. Would this option or anything similar still work ?
Thanks in advance