yiran-THU / MobileNet-YOLO

A caffe implementation of MobileNet-YOLO detection network

MobileNet-YOLO Caffe

MobileNet-YOLO

A caffe implementation of MobileNet-YOLO detection network , test on VOC2007

Network	mAP	Resolution	Download	NetScope	Inference time (GTX 1080)	Inference time (i5-4440)
MobileNet-YOLO-Lite	0.675	416	deploy	graph	N/A	N/A
MobileNet-YOLOv3-Lite	0.717	320	deploy	graph	6 ms	150 ms
MobileNet-YOLOv3-Lite	0.737	416	deploy	graph	11 ms	280 ms

Note : the yolo_detection_output_layer not be optimization , and the deploy model was made by merge_bn.py

Windows Version

Caffe-YOLOv2-Windows

Oringinal darknet-yolov3

Converter

test on coco_minival_lmdb (IOU 0.5)

Network	mAP	Resolution	Download	NetScope
yolov3	54.4	416	caffemodel	graph
yolov3-spp	59.3	608	caffemodel	graph

Performance

Compare with YOLO , (IOU 0.5)

Network	mAP	Weight size	Resolution	NetScope
MobileNet-YOLOv3-Lite	34.0*	21.5 mb	320	graph
MobileNet-YOLOv3-Lite	37.3*	21.5 mb	416	graph
MobileNet-YOLOv3	40.3*	22.5 mb	416	graph
YOLOv3-Tiny	33.1	33.8 mb	416

• (*) testdev-2015 server was closed , here use coco 2014 minival

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

CMake Build

Caffe page , dependency

> git clone https://github.com/eric612/MobileNet-YOLO.git 
> cd $MobileNet-YOLO_root/
> mkdir build
> cd build
> cmake ..
> make -j4
> make pycaffe

Training

Download lmdb

Unzip into $caffe_root/

Please check the path exist "$caffe_root\examples\VOC0712\VOC0712_trainval_lmdb" and "$caffe_root\examples\VOC0712\VOC0712_test_lmdb"

Download pre-trained weights , and save at $caffe_root\model\convert

> cd $caffe_root/
> sh train_yolo.sh

Demo

> cd $caffe_root/
> sh demo_yolo_lite.sh

If load success , you can see the image window like this

Future work

• Distillation training
• Openimages training and testing

License and Citation

Caffe is released under the BSD 2-Clause license. The BAIR/BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}