
摘要
迄今为止精度最高的目标检测器均基于由R-CNN提出的两阶段方法,该方法对一组稀疏的候选目标位置应用分类器。相比之下,针对可能目标位置进行规则且密集采样的单阶段检测器具有更快、更简洁的潜力,但迄今为止其精度仍落后于两阶段检测器。本文旨在探究造成这一差距的根本原因。我们发现,密集检测器在训练过程中遭遇的极端前景-背景类别不平衡是核心问题。为此,我们提出通过重构标准交叉熵损失函数来缓解类别不平衡问题,使模型对分类正确的样本赋予较低的损失权重。我们提出的新型Focal Loss能够将训练重点集中于少数困难样本,有效防止大量简单负样本在训练过程中淹没检测器。为验证该损失函数的有效性,我们设计并训练了一种简单的密集检测器,命名为RetinaNet。实验结果表明,当使用Focal Loss进行训练时,RetinaNet在保持与以往单阶段检测器相当速度的同时,其精度超越了所有现有的最先进两阶段检测器。代码地址:https://github.com/facebookresearch/Detectron。
代码仓库
trongnghia00/darknet
GitHub 中提及
fizyr/keras-retinanet
tf
GitHub 中提及
neshitov/UNet
pytorch
GitHub 中提及
vantupham/darknet
GitHub 中提及
yhenon/pytorch-retinanet
pytorch
GitHub 中提及
feidfoe/AdjustBnd4Imbalance
pytorch
GitHub 中提及
simonlevine/11785-project
pytorch
GitHub 中提及
unsky/focal-loss
mxnet
GitHub 中提及
mic-dkfz/medicaldetectiontoolkit
pytorch
GitHub 中提及
tianhai123/yolov3
GitHub 中提及
AlexeyAB/darknet
tf
GitHub 中提及
kyunghwan/darknet_v3
GitHub 中提及
anookeen/yolo
GitHub 中提及
daveboat/pytorch_focal_loss
pytorch
GitHub 中提及
Yashgh7076/Eluvio-ML-Scene-Segmentation
pytorch
GitHub 中提及
kornia/kornia
pytorch
witwickey/darknet
tf
GitHub 中提及
yiningzeng/darknet-fork-from-AlexeyAB
tf
GitHub 中提及
moallafatma/Breast_Cancer_Detection_Classification
pytorch
GitHub 中提及
Viveksbawa/SARAS-ESAD-Baseline
pytorch
GitHub 中提及
oorrppp2/darknet-pose
tf
GitHub 中提及
aakashjhawar/traffic-sign-detection
tf
GitHub 中提及
Guillem96/efficientdet-tf
tf
GitHub 中提及
teemoeric/projet
GitHub 中提及
DonCorle0ne/damage_detection
GitHub 中提及
IPLI/IPLI_Server
tf
GitHub 中提及
GXYM/Focal-loss
tf
GitHub 中提及
artemmavrin/focal-loss
tf
GitHub 中提及
cameronchoi/r3det-copy
pytorch
GitHub 中提及
toufiksk/darknet
GitHub 中提及
MariefW/6.-trainingYOLO
tf
GitHub 中提及
Shraddha2013/darknetyolo
tf
GitHub 中提及
stevenzhou2017/darknet_AlexeyAB
tf
GitHub 中提及
Popoooo/darknet_test
tf
GitHub 中提及
pytorch/vision
pytorch
samsonadmin/modified-alexeyab-darknet
tf
GitHub 中提及
artxtech/darknet-rnn
tf
GitHub 中提及
minhdua/PHONES
tf
GitHub 中提及
kongzhiyou/darknet-master
tf
GitHub 中提及
ryubidragonfire/mydarknet
GitHub 中提及
solapark/darknet_partdet
tf
GitHub 中提及
qingyujean/Classification-on-imbalanced-data
tf
GitHub 中提及
dishen12/py03
pytorch
GitHub 中提及
Wangxy2180/darknetKinectDetect
tf
GitHub 中提及
DInesh1234-wq/tyredataset
pytorch
GitHub 中提及
rnirdhar/yoloTestOneClass
GitHub 中提及
li-yibing/oceannet
GitHub 中提及
buptdbj/darknet-windows-linux
GitHub 中提及
jfc4050/detect-to-track
pytorch
GitHub 中提及
namdvt/Focal-loss-pytorch-implementation
pytorch
GitHub 中提及
saednasir/Helmet_Yolo
GitHub 中提及
Jo-dsa/SemanticSeg
pytorch
GitHub 中提及
RichardMathewsII/YOLBO
GitHub 中提及
simonlevine/clinical-longformer
pytorch
GitHub 中提及
gyes00205/waymo_tf_object_detection
tf
GitHub 中提及
MIC-DKFZ/RegRCNN
pytorch
GitHub 中提及
solapark/darknet_interpark
tf
GitHub 中提及
hweersot/darknet-custom
tf
GitHub 中提及
wangxianhao/Improved-YOLOv3-for-UAV
GitHub 中提及
williamccondori/YOLO-NFPA
GitHub 中提及
ermuur/darknet
GitHub 中提及
HongSic/DarknetAI
tf
GitHub 中提及
petrum01/Capstone_project_object_detection
tf
GitHub 中提及
skbadhsm/darknet2
tf
GitHub 中提及
NVIDIA/retinanet-examples
pytorch
GitHub 中提及
Shraddha2013/customyolo
tf
GitHub 中提及
rnirdhar/yoloMultiClass
GitHub 中提及
binhdv92/darknet_kriyeng
tf
GitHub 中提及
Shraddha2013/darknett
tf
GitHub 中提及
manishravula/yolov3_tiny_pruned
GitHub 中提及
YIZHE12/robots
GitHub 中提及
IlyaOvodov/AngelinaReader
pytorch
GitHub 中提及
zj463261929/darknet_mAP
GitHub 中提及
rebeccagio/object_detection
GitHub 中提及
hyang0129/foodclassapp
tf
GitHub 中提及
ghadahamed/darknet
tf
GitHub 中提及
solapark/darknet_partdet_tmp
tf
GitHub 中提及
CahideSara/deneme
tf
GitHub 中提及
kalelpark/ral
pytorch
GitHub 中提及
hankpark0706/darknet
GitHub 中提及
ghaniskn/GCorp-Darknet
GitHub 中提及
istvanmegyeri/audio_tagging
GitHub 中提及
patelmiteshn/darknet
GitHub 中提及
khaled2ahmed/k2a
tf
GitHub 中提及
lhcezx/Graph-FPN
tf
GitHub 中提及
zhongzisha/object_detection
tf
GitHub 中提及
aminekha/AI-For-2022
tf
GitHub 中提及
dc17540/darknet
tf
GitHub 中提及
travismyers19/Vulnerable_Buildings
tf
GitHub 中提及
huytranvan2010/RetinaNet-Understanding
GitHub 中提及
nguoido/Yolo-alexeyAB
tf
GitHub 中提及
ayushpalak/darknet_cokebottles
GitHub 中提及
darshans0200/YOLOTest
GitHub 中提及
yudie433/darknet
GitHub 中提及
Thinklab-SJTU/CSL_RetinaNet_Tensorflow
tf
GitHub 中提及
csm-kr/Retinanet_pytorch
pytorch
GitHub 中提及
soccergame/darknet
GitHub 中提及
tryolabs/luminoth
tf
GitHub 中提及
cloudnine148/PHC_2nd_SPC
GitHub 中提及
ruinmessi/RFBNet
pytorch
GitHub 中提及
open-mmlab/mmdetection
pytorch
DetectionTeamUCAS/RetinaNet_Tensorflow_Rotation
tf
GitHub 中提及
ivanwilliammd/I3DR-Net-Transfer-Learning
pytorch
GitHub 中提及
mathieuorhan/darknet
GitHub 中提及
SanketAgrawal/Catagorical-Focal-Loss
GitHub 中提及
xiamenwcy/extended-caffe
GitHub 中提及
xuanyuzhou98/SqueezeSeg
tf
GitHub 中提及
yytang2012/darknet
GitHub 中提及
fabiofumarola/ultrayolo
tf
GitHub 中提及
Maskify/darknet
tf
GitHub 中提及
rahowa/catboost_focal_loss
GitHub 中提及
kvijaykrish/Bosch_AI_VehicleClassification
tf
GitHub 中提及
facebookresearch/detectron2
pytorch
caihaocong/Yolov3_AlexeyAB_Version
tf
GitHub 中提及
tallysprado/myDarknet
tf
GitHub 中提及
joheras/darknet-colab
tf
GitHub 中提及
mkocabas/focal-loss-keras
tf
GitHub 中提及
tensorflow/models
tf
GitHub 中提及
raspstephan/sugar-flower-fish-or-gravel
GitHub 中提及
KingBoyBIT/yolov3test
GitHub 中提及
pierluigiferrari/ssd_keras
tf
GitHub 中提及
iskandari/darknet
tf
GitHub 中提及
Shraddha2013/mydarknetyolo
tf
GitHub 中提及
unsky/RetinaNet
mxnet
GitHub 中提及
RuoyuChen10/objectdetection-saliency-maps
pytorch
GitHub 中提及
Lmath11/darknet
tf
GitHub 中提及
2023-MindSpore-1/ms-code-44
mindspore
jolibrain/caffe
GitHub 中提及
sulasen/race-events-recognition-1
tf
GitHub 中提及
jndeng/DACSDC-DeepZ
GitHub 中提及
gokulprasadthekkel/pytorch-multi-class-focal-loss
pytorch
GitHub 中提及
EMsnap/RobotSorting
GitHub 中提及
wodyjowski/colab-training
tf
GitHub 中提及
ShaojieJiang/tldr
pytorch
GitHub 中提及
rainofmine/Face_Attention_Network
pytorch
GitHub 中提及
agutuyen-dev/darknet
tf
GitHub 中提及
AdeelH/pytorch-multi-class-focal-loss
pytorch
GitHub 中提及
edgarrt/custom_darknet
tf
GitHub 中提及
eric-erki/yolov3
GitHub 中提及
jiajunhua/facebookresearch-Detectron
caffe2
GitHub 中提及
Shraddha2013/myfileyolo
tf
GitHub 中提及
ShaharrHorn/Detection-Game-logic
GitHub 中提及
bharatsush/object_detection
GitHub 中提及
MIC-DKFZ/DetectionAndRegression
pytorch
GitHub 中提及
Shraddha2013/darknetyolo3
tf
GitHub 中提及
OFRIN/Tensorflow_RetinaFace
tf
GitHub 中提及
hhuaibo/darknet
tf
GitHub 中提及
Stephenfang51/Focal_loss_turtorial
pytorch
GitHub 中提及
vuongtrannguyenkhoi/darknet
GitHub 中提及
fznsakib/great-ape-behaviour-detector
pytorch
GitHub 中提及
viethungluu/OCRetina
tf
GitHub 中提及
ViswanathaReddyGajjala/EfficientNet-RetinaNet
pytorch
GitHub 中提及
JaryHuang/awesome_SSD_FPN_GIoU
pytorch
GitHub 中提及
DaloroAT/first_break_picking
pytorch
GitHub 中提及
Sushma07/dancedarknet
tf
GitHub 中提及
FL77N/RetinaNet-Based-on-PPdet
paddle
GitHub 中提及
Shraddha2013/darknetyol
tf
GitHub 中提及
lyk19940625/MyRFBNet
pytorch
GitHub 中提及
SJTU-Thinklab-Det/r3det-on-mmdetection
tf
GitHub 中提及
AtlasCoCo/Darknet
tf
GitHub 中提及
eric-erki/darknet
GitHub 中提及
liyemei/Deformable-active-convolutional
GitHub 中提及
hisiter97/darknet
tf
GitHub 中提及
e0015274/darknet
tf
GitHub 中提及
llsouder/screen-ocr
GitHub 中提及
yijiaceline/Final-Project-Group4
pytorch
GitHub 中提及
ChristianMarzahl/ObjectDetection
pytorch
GitHub 中提及
Guillem96/efficient-net-tf
tf
GitHub 中提及
sdu2011/darknet_alexyab
tf
GitHub 中提及
yuliani29/yolotraining
GitHub 中提及
ZTao-z/multiflow-resnet-ssd
pytorch
GitHub 中提及
sudharavali/objectDetectionYOLO
tf
GitHub 中提及
LIU1514/Yolov3-
tf
GitHub 中提及
YIZHE12/fashiontags
tf
GitHub 中提及
zzhuolun/IRL
pytorch
GitHub 中提及
Mind23-2/MindCode-176
mindspore
ayushpalak/darknet_multi_cokebottles
GitHub 中提及
shashigharti/secure-privateai-scholarship-challenge
pytorch
GitHub 中提及
497626895/darknet
tf
GitHub 中提及
dwaithe/darknet3AB
tf
GitHub 中提及
Suci2609/generate_train_list.py-
tf
GitHub 中提及
artynet/darknet-alexeyAB
tf
GitHub 中提及
benihime91/pytorch_retinanet
pytorch
GitHub 中提及
haithink/myWindowsDarknet
GitHub 中提及
facebookresearch/detectron
官方
pytorch
GitHub 中提及
pengrenlai/RetinaNet-tensorflow
tf
GitHub 中提及
magnificent1208/r3det-on-mmdetection
pytorch
GitHub 中提及
xytpai/retinanet
pytorch
GitHub 中提及
kuangliu/pytorch-fpn
pytorch
GitHub 中提及
binhdv92/darknet
tf
GitHub 中提及
detectionteamucas/retinanet_tensorflow
tf
GitHub 中提及
bobby20180331/darknet_pycharm
GitHub 中提及
dtthi/darknetAlexeyAB
tf
GitHub 中提及
Yonder-OSS/D3M-Primitives
tf
GitHub 中提及
saber2011/darknet
tf
GitHub 中提及
annabellachen/newDarknet
tf
GitHub 中提及
karan96/NewOne
tf
GitHub 中提及
edwardclem/deepscribe
pytorch
GitHub 中提及
ZhenJie-Zhang/darknet_shoe
tf
GitHub 中提及
kawshik8/DL-project
pytorch
GitHub 中提及
delmalih/MIAS-mammography-obj-detection
pytorch
GitHub 中提及
sidify/resnet_focal_loss
GitHub 中提及
ahhan02/darknet-alex
tf
GitHub 中提及
Techyee/darknet_resource
tf
GitHub 中提及
PolarisAlpha/darknet
GitHub 中提及
juergenlandauer/Maya-Challenge
pytorch
GitHub 中提及
solapark/da_yolo
tf
GitHub 中提及
DrMMZ/RetinaNet
tf
GitHub 中提及
mdv3101/darknet-yolov3
GitHub 中提及
hamzaMahdi/darknet
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 2d-object-detection-on-sardet-100k | RetinaNet | box mAP: 47.4 |
| dense-object-detection-on-sku-110k | RetinaNet | AP: 45.5 AP75: .389 |
| face-identification-on-trillion-pairs-dataset | F-Softmax | Accuracy: 39.80 |
| face-verification-on-trillion-pairs-dataset | F-Softmax | Accuracy: 37.14 |
| long-tail-learning-on-coco-mlt | Focal Loss(ResNet-50) | Average mAP: 49.46 |
| long-tail-learning-on-egtea | Focal loss (3D- ResNeXt101) | Average Precision: 59.09 Average Recall: 59.17 |
| long-tail-learning-on-voc-mlt | Focal Loss(ResNet-50) | Average mAP: 73.88 |
| object-counting-on-carpk | RetinaNet (2018) | MAE: 24.58 |
| object-detection-on-coco | RetinaNet (ResNet-101-FPN) | AP50: 59.1 AP75: 42.3 APL: 50.2 APM: 42.7 APS: 21.8 Hardware Burden: 4G Operations per network pass: box mAP: 39.1 |
| object-detection-on-coco | RetinaNet (ResNeXt-101-FPN) | AP50: 61.1 AP75: 44.1 APL: 51.2 APM: 44.2 APS: 24.1 Hardware Burden: 4G Operations per network pass: box mAP: 40.8 |
| object-detection-on-coco-o | RetinaNet (ResNet-50) | Average mAP: 16.6 Effective Robustness: 0.18 |
| pedestrian-detection-on-tju-ped-campus | RetinaNet | ALL (miss rate): 44.34 HO (miss rate): 71.31 R (miss rate): 34.73 R+HO (miss rate): 42.26 RS (miss rate): 82.99 |
| pedestrian-detection-on-tju-ped-traffic | RetinaNet | ALL (miss rate): 41.40 HO (miss rate): 61.60 R (miss rate): 23.89 R+HO (miss rate): 28.45 RS (miss rate): 37.92 |