Skip to content

MobileNetV2-YoloV3-Nano: 0.5BFlops 3MB HUAWEI P40: 6ms/img, YoloFace-500k:0.1Bflops 420KB:fire::fire::fire:

License

Notifications You must be signed in to change notification settings

dog-qiuqiu/MobileNet-Yolo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

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)

Network VOC mAP(0.5) COCO mAP(0.5) Resolution Inference time (NCNN/Kirin 990) Inference time (MNN arm82/Kirin 990) FLOPS Weight size
MobileNetV2-YOLOv3-Lite(our) 73.26 37.44 320 28.42 ms 18 ms 1.8BFlops 8.0MB
MobileNetV2-YOLOv3-Nano(our) 65.27 30.13 320 10.16 ms 5 ms 0.5BFlops 3.0MB
MobileNetV2-YOLOv3 70.7 & 352 32.15 ms & ms 2.44BFlops 14.4MB
MobileNet-SSD 72.7 & 300 26.37 ms & ms & BFlops 23.1MB
YOLOv5s & 56.2 416 150.5 ms & ms 13.2BFlops 28.1MB
YOLOv3-Tiny-Prn & 33.1 416 36.6 ms & ms 3.5BFlops 18.8MB
YOLOv4-Tiny & 40.2 416 44.6 ms & ms 6.9BFlops 23.1MB
YOLO-Nano 69.1 & 416 & ms & ms 4.57BFlops 4.0MB
  • Support mobile inference frameworks such as NCNN&MNN
  • The mnn benchmark only includes the forward inference time
  • The ncnn benchmark is the forward inference time + post-processing time(NMS...) of the convolution feature map.
  • Darknet Train Configuration: CUDA-version: 10010 (10020), cuDNN: 7.6.4,OpenCV version: 4 GPU:RTX2080ti

MobileNetV2-YOLOv3-Lite-COCO Test results

image

Application

Ultralight-SimplePose

image

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

Network Resolution Inference time (NCNN/Kirin 990) Inference time (MNN arm82/Kirin 990) FLOPS Weight size
UltraFace-version-RFB 320x240 &ms 3.36ms 0.1BFlops 1.3MB
UltraFace-version-Slim 320x240 &ms 3.06ms 0.1BFlops 1.2MB
yoloface-500k 320x256 5.5ms 2.4ms 0.1BFlops 0.52MB
yoloface-500k-v2 352x288 4.7ms &ms 0.1BFlops 0.42MB
  • 都500k了,要啥mAP:sunglasses:
  • Inference time (DarkNet/i7-6700):13ms
  • The mnn benchmark only includes the forward inference time
  • The ncnn benchmark is the forward inference time + post-processing time(NMS...) of the convolution feature map.

Wider Face Val

Model Easy Set Medium Set Hard Set
libfacedetection v1(caffe) 0.65 0.5 0.233
libfacedetection v2(caffe) 0.714 0.585 0.306
Retinaface-Mobilenet-0.25 (Mxnet) 0.745 0.553 0.232
version-slim-320 0.77 0.671 0.395
version-RFB-320 0.787 0.698 0.438
yoloface-500k-320 0.728 0.682 0.431
yoloface-500k-352-v2 0.768 0.729 0.490
  • yoloface-500k-v2:The SE&CSP module is added
  • V2 does not support MNN temporarily
  • wider_face_val(ap05): yoloface-500k: 53.75 yoloface-500k-v2: 56.69

YoloFace-500k Test results(thresh 0.7)

image

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

image

YoloFace-50k: Sub-millisecond face detection model

Network Resolution Inference time (NCNN/Kirin 990) Inference time (MNN arm82/Kirin 990) Inference time (DarkNet/R3-3100) FLOPS Weight size
yoloface-50k 56x56 0.27ms 0.31ms 0.5 ms 0.001BFlops 46kb
  • For the close-range face detection model in a specific scene, the recommended detection distance is 1.5m

YoloFace-50k Test results(thresh 0.7)

image

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

Network Resolution Inference time (NCNN/Kirin 990) Inference time (MNN arm82/Kirin 990) Weight size
landmark106 112x112 0.6ms 0.5ms 1.4MB
  • Face detection: yoloface-50k Landmark: landmark106

YoloFace50k-landmark106 Test results

image

Reference&Framework instructions&How to Train

  • https://github.com/AlexeyAB/darknet
  • You must use a pre-trained model to train your own data set. You can make a pre-trained model based on the weights of COCO training in this project to initialize the network parameters
  • 交流qq群:1062122604

About model selection

  • MobileNetV2-YOLOv3-SPP: Nvidia Jeston, Intel Movidius, TensorRT,NPU,OPENVINO...High-performance embedded side
  • MobileNetV2-YOLOv3-Lite: High Performance ARM-CPU,Qualcomm Adreno GPU, ARM82...High-performance mobile
  • MobileNetV2-YOLOv3-NANO: ARM-CPU...Computing resources are limited
  • MobileNetV2-YOLOv3-Fastest: ....... Can you do personal face detection???It’s better than nothing

NCNN conversion tutorial

NCNN C++ Sample

NCNN Android Sample

image

DarkNet2Caffe tutorial

Environmental requirements

  • Python2.7
  • python-opencv
  • Caffe(add upsample layer https://github.com/dog-qiuqiu/caffe)
  • You have to compile cpu version of caffe!!!
      cd darknet2caffe/
      python darknet2caffe.py MobileNetV2-YOLOv3-Nano-voc.cfg MobileNetV2-YOLOv3-Nano-voc.weights MobileNetV2-YOLOv3-Nano-voc.prototxt MobileNetV2-YOLOv3-Nano-voc.caffemodel
      cp MobileNetV2-YOLOv3-Nano-voc.prototxt sample
      cp MobileNetV2-YOLOv3-Nano-voc.caffemodel sample
      cd sample
      python detector.py
    

MNN conversion tutorial

  • Benchmark:https://www.yuque.com/mnn/cn/tool_benchmark
  • Convert darknet model to caffemodel through darknet2caffe
  • Manually replace the upsample layer in prototxt with the interp layer
  • Take the modification of MobileNetV2-YOLOv3-Nano-voc.prototxt as an example
	#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