mAP (mean Average Precision) for Object Detection, Medium article by Johnathan Hui
The video above, entitled “Jetson Nano TensorRT FP16 YOLOv4x Mish 640”, can be found on Karol Majek's YouTube channel. Karol always tries out all possible object detectors.
The example above supposedly shows the TensorRT implementation / configuration yolov4x-mish-640 (FP16) by JK Jung. The Git Repo by JK Jung: https://github.com/jkjung-avt/tensorrt_demos#yolov4
Below he lists benchmarks for various detectors on the Jetson Nano in a rather large table. What I don't understand: JK Jung gives 1.46 FPS, in Karol's video 0.7 FPS are displayed at the top (about one second). Why is the video still running smoothly and also the bounding box updates are, in my opinion, much faster than 1 Hz?
The DNN to be used can be set with
detectnet --network=...
. The default is ssd-mobilet-v2. This was trained with 91 object classes from the COCO data set (labeled images).
The 91 known objects the system is trained to detect is listet here.
This video is much more plausible for me:
In the following video, Dustin Franklin explains how to train your own object classes:
However, the objects are extremely easy to discriminate based on their color alone.