abrenjy / mobilenet-yolo

A caffe implementation of MobileNet-YOLO detection network

MobileNet-YOLO Caffe

MobileNet-YOLO

A caffe implementation of MobileNet-YOLO detection network , first train on COCO trainval35k then fine-tune on 07+12 , test on VOC2007

Network	mAP	Resolution	Download	NetScope	Inference time (GTX 1080)	Inference time (i5-4440)
MobileNet-YOLOv3-Lite	74.6	320	caffemodel	graph	4.79 ms	150 ms
MobileNet-YOLOv3-Lite	76.3	416	caffemodel	graph	6.52 ms	280 ms

• inference time was log from script , does not include pre-processing
• the benchmark of cpu performance on Tencent/ncnn framework
• the deploy model was made by merge_bn.py , or you can try my custom version
• bn_model download here

Coverted tensorRT models

TensorRT-Yolov3-models

Knowledge Transfer

I use the following training path to improve accuracy , and decrease lite version trainning time

• First , train MobileNet-YOLOv3 on coco dataset (IOU_0.5 : 40.2 mAP)
• Second , train MobileNet-YOLOv3-Lite on coco dataset , pretrain weights use the first step output (IOU_0.5 : 38.9 mAP)
• Finally , train MobileNet-YOLOv3-Lite on voc dataset , pretrain weights use the second step output (76.3 mAP)

Windows Version

Caffe-YOLOv3-Windows

Oringinal darknet-yolov3

Converter

test on coco_minival_lmdb (IOU 0.5)

Network	mAP	Resolution	Download	NetScope
yolov3	54.2	416	caffemodel	graph
yolov3-spp	59.8	608	caffemodel	graph

• I haven't implement correct_yolo_boxes and relative function , so here only support square input resolution

Performance

Train on COCO trainval35k (2014) , and compare with YOLO , (IOU 0.5)

Network	IOU 0.5:0.95	IOU 0.5	IOU 0.75	Weight size	Resolution	NetScope	Resize Mode
MobileNet-YOLOv3-Lite	19.9	35.5	19.6	22.0 mb	320	graph	WARP
MobileNet-YOLOv3-Lite	21.5	38.9	21.2	22.0 mb	416	graph	WARP
MobileNet-YOLOv3	22.7	40.2	22.6	22.5 mb	416	graph	LetterBox
YOLOv3-Tiny		33.1		33.8 mb	416
MobileNet-YOLOv3-Lite-trt		37.5		23.5 mb	416	graph	WARP

• (*) testdev-2015 server was closed , here use coco 2014 minival
• MobileNet-YOLOv3-Lite-trt was the fastest model

Other Models

You can find non-depthwise convolution network here , Yolo-Model-Zoo

network	mAP	resolution	macc	param
PVA-YOLOv3	0.703	416	2.55G	4.72M
Pelee-YOLOv3	0.703	416	4.25G	3.85M

Model visulization tool

Supported on Netron , browser version

Build , Run and Training

See wiki

License and Citation

Please cite MobileNet-YOLO in your publications if it helps your research:

@article{MobileNet-YOLO,
  Author = {eric612 , Avisonic , ELAN},
  Year = {2018}
}

Reference

https://github.com/weiliu89/caffe/tree/ssd

https://pjreddie.com/darknet/yolo/

https://github.com/gklz1982/caffe-yolov2

https://github.com/yonghenglh6/DepthwiseConvolution

https://github.com/alexgkendall/caffe-segnet

https://github.com/BVLC/caffe/pull/6384/commits/4d2400e7ae692b25f034f02ff8e8cd3621725f5c

Cudnn convolution

https://github.com/chuanqi305/MobileNetv2-SSDLite/tree/master/src