Exploring varied object detection architectures

* Videos have been sped up so that it is not so boring! I set the faster clip to 30 FPS and the other at 9.4 FPS, so the proportional speed still maintains.

When deploying machine learning models to the real world, specially in resource constrained environments, model size and complexity plays a big role into determining an application’s usability, i.e.: will it make predictions that are correct and also fast enough?

In order to have a general idea of your hardware impact on processing time for object detection, I’ve made a preliminary benchmark table on a set of 1600 JPEG frames with 1280x720 resolution on the following hardware:

Hardware	CPU	RAM	Graphics card
Raspberry Pi 3 B+	1.4GHz 64-bit quad-core ARM Cortex-A53 CPU	1GB LPDDR2 SDRAM	None
Raspberry Pi 3 B+	1.4GHz 64-bit quad-core ARM Cortex-A53 CPU	1GB LPDDR2 SDRAM	Movidius Neural Compute Stick (not a video card, but helps for GPU processing)
Dell G3 15	Intel(R) Core(TM) i5-8300H CPU @ 2.30GHz	8GB SODIMM DDR4 Synchronous	GeForce GTX 1050 Mobile

Making inferences on top of a MobileNet with SSD architecture and weights totalling 22.1 MB:

Hardware	FPS	Time per frame (s)
Raspberry Pi 3 B+ on CPU	1.479	0.676
Raspberry Pi 3 B+ with Movidius NCU	4.721	0.212
Dell G3 15 on GPU	(TODO)	(TODO)