Home

Awesome

2021.2.6 此项目不再更新,新项目地址: Yolo-Fastest: Faster and stronger https://github.com/dog-qiuqiu/Yolo-Fastest

image

***Darknet Group convolution is not well supported on some GPUs such as NVIDIA PASCAL!!!

针对某些Pascal显卡例如1080ti在darknet上 训练失败/训练异常缓慢/推理速度异常 的可以采用Pytorch版yolo3框架 训练/推理

MobileNetV2-YOLOv3-Lite&Nano Darknet

Mobile inference frameworks benchmark (4*ARM_CPU)

NetworkVOC mAP(0.5)COCO mAP(0.5)ResolutionInference time (NCNN/Kirin 990)Inference time (MNN arm82/Kirin 990)FLOPSWeight size
MobileNetV2-YOLOv3-Lite(our)73.2637.4432028.42 ms18 ms1.8BFlops8.0MB
MobileNetV2-YOLOv3-Nano(our)65.2730.1332010.16 ms5 ms0.5BFlops3.0MB
MobileNetV2-YOLOv370.7&35232.15 ms& ms2.44BFlops14.4MB
MobileNet-SSD72.7&30026.37 ms& ms& BFlops23.1MB
YOLOv5s&56.2416150.5 ms& ms13.2BFlops28.1MB
YOLOv3-Tiny-Prn&33.141636.6 ms& ms3.5BFlops18.8MB
YOLOv4-Tiny&40.241644.6 ms& ms6.9BFlops23.1MB
YOLO-Nano69.1&416& ms& ms4.57BFlops4.0MB

MobileNetV2-YOLOv3-Lite-COCO Test results

image

Application

Ultralight-SimplePose

image

YoloFace-500k: 500kb yolo-Face-Detection

NetworkResolutionInference time (NCNN/Kirin 990)Inference time (MNN arm82/Kirin 990)FLOPSWeight size
UltraFace-version-RFB320x240&ms3.36ms0.1BFlops1.3MB
UltraFace-version-Slim320x240&ms3.06ms0.1BFlops1.2MB
yoloface-500k320x2565.5ms2.4ms0.1BFlops0.52MB
yoloface-500k-v2352x2884.7ms&ms0.1BFlops0.42MB

Wider Face Val

ModelEasy SetMedium SetHard Set
libfacedetection v1(caffe)0.650.50.233
libfacedetection v2(caffe)0.7140.5850.306
Retinaface-Mobilenet-0.25 (Mxnet)0.7450.5530.232
version-slim-3200.770.6710.395
version-RFB-3200.7870.6980.438
yoloface-500k-3200.7280.6820.431
yoloface-500k-352-v20.7680.7290.490

YoloFace-500k Test results(thresh 0.7)

image

YoloFace-500k-v2 Test results(thresh 0.7)

image

YoloFace-50k: Sub-millisecond face detection model

NetworkResolutionInference time (NCNN/Kirin 990)Inference time (MNN arm82/Kirin 990)Inference time (DarkNet/R3-3100)FLOPSWeight size
yoloface-50k56x560.27ms0.31ms0.5 ms0.001BFlops46kb

YoloFace-50k Test results(thresh 0.7)

image

YoloFace50k-landmark106(Ultra lightweight 106 point face-landmark model)

NetworkResolutionInference time (NCNN/Kirin 990)Inference time (MNN arm82/Kirin 990)Weight size
landmark106112x1120.6ms0.5ms1.4MB

YoloFace50k-landmark106 Test results

image

Reference&Framework instructions&How to Train

About model selection

NCNN conversion tutorial

NCNN C++ Sample

NCNN Android Sample

image

DarkNet2Caffe tutorial

Environmental requirements

MNN conversion tutorial

	#layer {
	#    bottom: "layer71-route"
	#    top: "layer72-upsample"
	#    name: "layer72-upsample"
	#    type: "Upsample"
	#    upsample_param {
	#        scale: 2
	#    }
	#}
	layer {
	    bottom: "layer71-route"
	    top: "layer72-upsample"
	    name: "layer72-upsample"
	    type: "Interp"
	    interp_param {
		height:20  #upsample h size
		width:20   #upsample w size
	    }
	}

Thanks