MMDetection 톺아보기

MMDetection

(object detection tool box and benchmark)

• MMDetection Paper : Here
• Official code : Here

object detection tool box인 MMDetection과 MMDetection이 지원하는 프레임워크들의 benchmark를 알아보자

Frameworks

지원하는 프레임워크 KeyPoint

Single stage

input -> feature extraction -> detection(Localization, Classification ) -> output(multi class classification,bounding box regression)

Localization, Classification 을 동시에 해결

Name	Content	Year
SSD	multi scale feature map	2015
RetinaNet	Focal loss	2017
GHM	gradient harmonizing mechanism	2019
FCOS	fully convolutional ,anchor-free	2019
FSAF	fully convolutional ,anchor-free	2019

Two stage

input -> region proposal(Localization) -> Classification -> output(multi class classification,bounding box regression)

Localization, Classification 을 순차적으로 해결

Name	Content	Year
Fast R-CNN	Region Proposal(RP) , ROI Pooling	2015
Faster R-CNN	Region Proposal Network(RPN),Fast R-CNN	2015
R-FCN	fully convolutional, Faster R-CNN	2016
Mask R-CNN	Binary Mask ,RoI Align , Faster RCNN	2017
Grid R-CNN	grid guided localization mechanism(bounding box regression imporved), RPN	2018
Mask Scoring R-CNN	mask IoU prediction, Mask R-CNN	2019
Double-Head R-CNN	convolution head(localization) + fully connected head(classification)	2019

Multi Stage

Name	Content	Year
Casecade R-CNN	multi-stage	2017
Hybrid Task Cascade	multi-stage , multi-branch , instance segmentation	2019

General Modules and Methods

Name	Content	Year
Mixed Precision Training	half precision floating point (FP16)	2018
Soft NMS	new NMS	2017
OHEM	hard sampling	2016
DCN	deformable convolution, deformable RoI pooling	2017
DCNv2	deformable operators	2018
ScratchDet	scratch,random initialization	2018
Train from Scratch	scratch	2018
M2Det	effective feature pyramids	2018
GCNet	global context block	2019
Generalized Attention	generalized attention formulation	2019
SyncBN,MegDet	batch normalization, synchronized	2017
GroupNormalization	group batch normalization	2018
Weight Standardization	micro-batch training	2019
HRNet	high-resolution representations, backbone	2019
Guided Anchoring	new anchoring, sparse and arbitrary-shaped anchors	2019
Libra R-CNN	framework, balanced learning	2019

Architecture

Model Representations

• Backbone : fully connected layer가 없는 resnet-50

• Neck : feature map 수정/재구성 ex) FPN

• DenseHead : AnchorHead / AnchorFreeHead(RPNHead, RetinaHead, FCOSHead)를 포함하고 feature map의 밀집된 위치에서 작동한다.

• RoIExtractor : RoIPooling과 같은 연산을 사용해 ROIwise feature를 추출하는 부분이다. ex) SingleRoI

• RoIHead : bounding box를 분류, 회귀, 마스크 예측

Training Pipeline

• hooking : 함수 호출, 메시지, 이벤트 등을 중간에서 바꾸거나 가로채는 명령, 방법, 기술이나 행위를 말한다.

hooking mechanism을 가지고 있는 pipeline

Benchmark

Dataset

• VOC
• COCO

부록 : 사용하기

참조 : 깃허브

환경 설치하기

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab

conda install pytorch torchvision -c pytorch

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection

pip install mmcv

python setup.py develop
# or "pip install -v -e ."

mkdir data
ln -s $COCO_ROOT data

데이터셋 준비하기

mmdetection
├── mmdet
├── tools
├── configs
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017
│   ├── cityscapes
│   │   ├── annotations
│   │   ├── train
│   │   ├── val
│   ├── VOCdevkit
│   │   ├── VOC2007
│   │   ├── VOC2012

데이터셋을 다운로드하고 위와 같은 구조를 맞추어 주어야한다.

• COCO
• VOC2007
• VOC2012
• cityscapes

cd data/cityscapes/
mv train/*/* train/

모델 준비하기

다운로드 : https://github.com/open-mmlab/mmdetection/blob/master/docs/MODEL_ZOO.md

• 실행하기

dataset demo

python tools/test.py configs/faster_rcnn_r50_fpn_1x.py \
    checkpoints/faster_rcnn_r50_fpn_1x_20181010-3d1b3351.pth \
    --show

bbox, mask AP

python tools/test.py configs/mask_rcnn_r50_fpn_1x.py \
    checkpoints/mask_rcnn_r50_fpn_1x_20181010-069fa190.pth \
    --out results.pkl --eval bbox segm

webcam demo

python demo/webcam_demo.py configs/faster_rcnn_r50_fpn_1x.py \
    checkpoints/faster_rcnn_r50_fpn_1x_20181010-3d1b3351.pth

이것저것 테스팅을 해볼수 있는 유용한 toolbox다. 고성능 API도 이용할 수 있기 때문에 사용이 편리하다.