Q-engineering

@dipu0, If you have an 64-OS, you can use the same code. Of course, the inference times are longer.

Just done. It works like a charm.

You can use a Raspberry Pi 3B+, or the new Zero 2 W. It is even possible on a 'old' Zero, however with its simple processor (ARMv6 - without NEON)...

It all depends on the quantization tool of the ncnn framework. As you no doubt know, deep learning models consist of many layers with different operations. As time goes by,...

Keep in mind that exporting Ultralytics YoloV5 or YoloV8 can be an adventure when it comes to ncnn. See: https://github.com/Tencent/ncnn/discussions/4541 Especially the already quantized version are notorious issue generators.

1) Try quantization, just as you suggested. Keep in mind that it can fail. 2) Try overclocking (simple and works always). 3) If possible reduce input size 320x320 is faster...

Which system are you comparing? In general, the answer is that it depends on your framework. With dedicated CUDA frameworks, like TensorRT, CUDA is faster. On the other hand, using...

I don't have any experiments. Finding a model suitable for TensorRT is most of the time the difficult part. There are a lot of models out there, but building a...

Dear @rsingh2083, How do you want to prioritize the found objects? For instance, you found 8 objects. Which two do you want to show? The two with the highest probability?...

Start playing with cv::Rect. Your outcome has a rect around the person, and if you define the 4 regions as rect, you can find the overlap. See https://answers.opencv.org/question/67091/how-to-find-if-2-rectangles-are-overlapping-each-other/ for the...