目标检测汇总(随心所欲更新)
梳理一遍,还需补充哪些欢迎指出。 指标 MS COCO
一、单阶段目标检测(基于anchor的)
multi-scale | backbone | arxiv/ github | AP | AP50 | AP75 | APs | APM | APL | 亮点 | time | |
YOLO | DarkNet-53 | ||||||||||
SSD300(Google) |
VGG | https://arxiv.org/pdf/1512.02325.pdf | 23.2 | 41.2 | 23.4 | 5.3 | 23.2 | 38.6 | |||
SSD512x512 | VGG | 26.8 | 46.5 | 27.8 | 9.0 | 28.9 | 41.9 | ||||
RetinaNet (FAIR facebook) |
ResNet101FPN | https://arxiv.org/pdf/1708.02002.pdf | 39.1 | 59.1 | 42.3 | 21.8 | 42.7 | 50.2 | Focal loss 类别平衡损失 | ||
ResNeXt101 | 40.8 | 61.1 | 44.1 | 24.1 | 44.2 | 51.2 | |||||
RefineDet512 ( 中科院) |
ResNet101 | https://arxiv.org/pdf/1711.06897.pdf | 36.4 | 57.5 | 39.5 | 16.6 | 39.9 | 51.4 | |||
√ | 41.8 | 62.9 | 45.7 | 25.6 | 45.1 | 54.1 | |||||
M2Det (阿里达摩+北大) |
VGG16 | https://arxiv.org/pdf/1811.04533.pdf | 41.0 | 59.7 | 45.0 | 22.1 | 46.5 | 53.8 | |||
512x512 | √ | VGG16 | 44.2 | 64.6 | 49.3 | 29.2 | 47.9 | 55.1 | |||
800x800 | ResNet101 | 38.8 | 59.4 | 41.7 | 20.5 | 43.9 | 53.4 | ||||
√ | ResNet101 | 43.9 | 64.4 | 48.4 | 29.6 | 49.6 | 54.3 | ||||
YOLO-v4 512X512 | CSPDarknet-53 | https://arxiv.org/pdf/2004.10934.pdf | 43.0 | 64.9 | 46.5 | 24.3 | 46.1 | 55.2 | |||
608 | CSPDarknet-53 | 43.5 | 65.7 | 47.3 | 26.7 | 46.7 | 53.3 |
二、Anchor-Free
|
multi-scale | arxiv | backbone | AP | AP50 | AP75 | APS | APM | APL | 亮点 |
CornetNet | https://arxiv.org/pdf/1808.01244.pdf | Hourglass 104 | 40.5 | 56.5 | 43.1 | 19.4 | 42.7 | 53.9 | ||
√ | Hourglass 104 | 42.1 | 57.8 | 45.3 | 20.8 | 44.8 | 56.7 | |||
ExtremNet | https://arxiv.org/pdf/1901.08043.pdf | Hourglass 104 | 40.2 | 55.5 | 43.2 | 20.4 | 43.2 | 53.1 | ||
√ | Hourglass 104 | 43.7 | 60.5 | 47.0 | 24.1 | 46.9 | 57.6 | |||
FSAF800 | https://arxiv.org/pdf/1903.00621.pdf | ResNext-101 | 42.9 | 63.8 | 46.3 | 26.6 | 46.2 | 52.7 | ||
FSAF | √ | 44.6 | 65.2 | 48.6 | 29.7 | 47.1 | 54.6 | |||
FCOS | https://arxiv.org/pdf/1904.01355.pdf | Hourglass109 | 40.5 | 56.5 | 43.1 | 19.4 | 42.7 | 53.9 | ||
ResNeXt-64x4d-101-FPN | 44.7 | 64.1 | 48.1 | 27.6 | 47.5 | 55.6 | ||||
FoveaBox | https://arxiv.org/pdf/1904.03797v1.pdf | resnet101 | 40.6 | 60.1 | 43.5 | 23.3 | 45.2 | 54.5 | ||
ResNeXt101 | 42.1 | 61.9 | 45.2 | 24.9 | 46.8 | 55.6 | ||||
RPDet | https://arxiv.org/pdf/1904.11490.pdf | ResNet101 | 41.0 | 62.9 | 44.3 | 23.6 | 44.1 | 51.7 | ||
√ | ResNet101-DCN | 43.5 | 61.3 | 46.7 | 25.3 | 45.3 | 55.0 | |||
CenterNet:keypoint triplet | https://arxiv.org/pdf/1904.08189.pdf | Hourglass 52 | 41.6 | 59.4 | 44.2 | 22.5 | 43.1 | 54.1 | ||
Hourglass 104 | 44.9 | 62.9 | 48.1 | 25.6 | 47.4 | 57.4 | ||||
√ | Hourglass 52 | 43.5 | 61.3 | 46.7 | 25.3 | 45.3 | 55.0 | |||
√ | Hourglass 104 | 47.0 | 64.5 | 50.7 | 28.9 | 49.9 | 58.9 |
三、二阶段目标检测:
muti-scale | backbone | arxiv | AP | AP50 | AP75 | APS | APM | APL | |||
RCNN | |||||||||||
Fast | |||||||||||
Faster | |||||||||||
FPN | https://arxiv.org/pdf/1612.03144.pdf | ||||||||||
Mask r-cnn | https://arxiv.org/pdf/1703.06870.pdf | 39.8 | 62.3 | 43.4 | 22.1 | 43.2 | 51.2 | ||||
IoUNet | resnet101 | https://arxiv.org/pdf/1807.11590.pdf | 40.6 | 59.0 | - | - | - | - | |||
Libra-RCNN | resnext101 | https://arxiv.org/pdf/1904.02701.pdf | 43.0 | 64 | 47.0 | 25.3 | 45.6 | 54.6 | |||
DetNet | DetNet59 | https://arxiv.org/pdf/1804.06215.pdf | 40.2 | 61.7 | 43.7 | 23.9 | 43.2 | 52.0 | |||
R-DAD | resnet101 | https://arxiv.org/pdf/1901.08225.pdf | 40.4 | 60.5 | 43.7 | 20.4 | 45 | 56.1 | |||
√ | resnet101 | 43.1 | 63.5 | 47.4 | 24.1 | 45.9 | 54.7 | ||||
Cacscade RCNN | resnet101 | https://arxiv.org/abs/1712.00726 | 42.8 | 62.1 | 46.3 | 23.7 | 45.5 | 55.2 | |||
EfficientDet | https://arxiv.org/pdf/1911.09070.pdf | 52.2 | 71.4 | 56.3 |
RCNN:(1)用selective search选取候选区, CNN提取特征 SVM分类 (2) 对不同的候选区都要单独用CNN提特征造成耗时
Fast-RCNN:(1) 提取候选框,对原图直接CNN提取特征,将候选框映射到特征图上。
Faster-RCNN:(1)提出RPN (2)实现端到端训练
FPN (CVPR2017) :
Mask-RCNN:(1)ROI Align 替换了 ROI pooling。 (2)roi Align后的特增加一个分割掩码的分支
DetNet:
IoUNet:(1)precisePooling ( 2 ) 增加了新的分支回归候选区的IoU大小,提出根据IoU指导的NMS。
R-DAD:
Libra-RCNN:(1)FPN改进:rescale strengthen:和原始特征融合(2)正负样本采样:正样本按类别均衡采样,负样本使用IoU间隔法采样(3)回归损失函数:smooth L1 => Balanced smooth L1
Cacscade RCNN:(1)使用了三个分支回归和分类,下一个分支都使用上一个分支回归的框作为输入。不同分支的正样本使用不同的阈值划分。
四、轻量化的目标检测
input size | backbone | arxiv | AP | AP50 | AP75 | 亮点 | ||
ThunderNet | 320x320 | SNet146 | https://arxiv.org/pdf/1903.11752.pdf | 23.6 | 40.2 | 24.5 | ||
320x320 | SENet535 | 28.0 | 46.2 | 29.5 | ||||