Why are most models either light or heavy?

[harsco-jfernandez]

There are heavy/large models like yolov3 and yolov4 and light/small models like yolov3-tiny. These give you higher mAP + low FPS or lower mAP + high FPS.

Why is it that there aren't any models that are balanced(goldie lock zone, not too hot not too cold)? Or are there any?

[AlexeyAB]

yolov4.cfg/weights (416x416 batch=1) + tkDNN-TensorRT + Jetson Xavier AGX = 32 FPS (realtime) - 43.5% AP - 65.7% AP50 - is balanced goldie lock zone ... that lies on Pareto optimal curve. The best ever accuracy with the lowest possible latency for Real-time systems on embedded device.

[harsco-jfernandez]

Thank you @AlexeyAB.

I may use that in the future in a more complex dataset. yolo_v3_tiny_pan3_aa_ae_mixup is giving me an excellent mAP in the small dataset I'm using now. I will attempt to get tiny_pan3_aa to run on tkDNN, which I'm liking - deeepstream's gstreamer is overly complex.

Thanks again, I appreciate your effort. I fear someone will take notice of you any day and will steel you away from us. ;)

[AlexeyAB]

You can try to use OpenCV (compiled with CUDA+cuDNN) without Darknet for detection: https://docs.opencv.org/master/da/d9d/tutorial_dnn_yolo.html

OpenCV is only slightly slower than tkDNN-TensorRT: #5354 (comment)

How to use:

• python: https://github.com/opencv/opencv/blob/master/samples/dnn/object_detection.py
• c++: https://github.com/opencv/opencv/blob/master/samples/dnn/object_detection.cpp

[harsco-jfernandez]

@AlexeyAB

Awesome... I'll give it a go...

You know what else would be really nice... yolov4-tiny ;) ;)

Enjoy your week and thanks again

[vinorth-v]

I would also like to know if it is expected to have yolov4 tiny?