Skip to content

Latest commit

 

History

History
82 lines (60 loc) · 2.64 KB

README.md

File metadata and controls

82 lines (60 loc) · 2.64 KB
Raw

SSD_ResNet_Pytorch

SSD: Single Shot MultiBox Detector

This is a repositories that replaced the primary network(VGG) in an SSD paper with Resnet.

Getting Started

Download git and dataset

git clone https://github.com/cjf8899/SSD_ResNet_Pytorch.git

cd SSD_ResNet_Pytorch

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
tar xvf VOCtrainval_06-Nov-2007.tar
mv ./VOCdevkit ./VOCtrainval2007

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
tar xvf VOCtrainval_11-May-2012.tar
mv ./VOCdevkit ./VOCtrainval2012

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
tar xvf VOCtest_06-Nov-2007.tar
mv ./VOCdevkit ./VOCtest2007

the structures would like

~/VOCtrainval2007/
    -- VOC2007
~/VOCtrainval2012/
    -- VOC2012  
~/VOCtest2007/
    -- VOC2007

Train

pre-train VGG weights (https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth)

pre-train ResNet18 weights (https://download.pytorch.org/models/resnet18-5c106cde.pth)

pre-train ResNet101 weights (https://download.pytorch.org/models/resnet101-5d3b4d8f.pth)

mkdir weights
cd weights
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
wget https://download.pytorch.org/models/resnet18-5c106cde.pth
wget https://download.pytorch.org/models/resnet101-5d3b4d8f.pth

Select and use the desired model

Results

The DSSD paper measured the performance of resnet101. The image size of the DSSD paper is 321.

Implementation mAP
SSD paper - VGG 77.2
DSSD paper - ResNet101 77.1
this repo - VGG 78.9
this repo - ResNet18 70.1
this repo - ResNet101 77.5

The SSD paper used Xavier after the base network(VGG), but did not use Xavier because this repo used relu.

Other experiments..

Implementation mAP
ResNet18_No_pretrain 54.4
VGG_No_Augmentation 66.8

Paper Review

VGG PR : https://cjf8899.github.io/image%20recognition/VGG/

ResNet PR : https://cjf8899.github.io/image%20recognition/ResNet/

ResNeXt PR : https://cjf8899.github.io/image%20recognition/ResNeXt/

Referenced. Thank you all

ssd code : https://github.com/HosinPrime/simple-ssd-for-beginners