YOLOV4论文解读与应用
YOLOV4
目标检测相关博客查看: CV综述目标检测整理—目录
论文地址:
YOLOv4: Optimal Speed and Accuracy of Object Detection
源码地址:
- https://github.com/AlexeyAB/darknet
- pytorch: Tianxiaomo-pytorch-YOLOv4
- pytorch2: yuto3o-yolox
- pytorch3: WongKinYiu-PyTorch_YOLOv4
训练TODO##############
camera测试
(torchpy36) air@air-System-Product-Name:~/model-train/yolo/pytorch-YOLOv4$ python demo.py -cfgfile cfg/yolov4.cfg -weightfile ./checkpoints/yolov4.weights
参考博客:
如何评价新出的YOLO v4 ?
yolo系列之yolo v3【深度解析】
【深度学习MobileNet】——深刻解读MobileNet网络结构
模型特点:
- trick万花筒
- Bag of freebies
- Bag of Specials;
- YOLOv4的创新点
优点
- 深度可分离卷积depthwise-conv 代替普通卷积
- pointwise conv
- 共28层
缺点
- kernel容易训练废掉(Relu激活函数):
参考博客:轻量级模型:MobileNet V2
yolov3-spp
论文地址:
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications @20190321
源码地址: darknet
pytorch: ultralytics-yolov3
需要apex,提供混合精度模型训练,降低训练所需要的GPU资源
nvidia-apex apex
训练yolov3-spp: coco2017数据
(torchpy36) air@air-System-Product-Name:~/model-train/yolo/yolov3/ultralytics-yolov3$ python train.py
参考博客:
轻量级模型:MobileNet V2
MobileNet V2 论文初读
模型特点:
- Linear Bottleneck
- Inverted Residuals
- relu6
- 54层
*****
论文地址:
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications @20170417
参考博客:
【深度学习MobileNet】——深刻解读MobileNet网络结构
模型特点:
- 深度可分离卷积depthwise-conv 代替普通卷积
- pointwise conv
参考博客
深度学习MobileNet——深刻解读MobileNet网络结构
MobileNet v1、v2
Focal Loss for Dense Object Detection解读
RetinaNet: Focal loss在目标检测网络中的应用
人脸检测:RetinaFace(开源简化版)详细解读
论文解读
论文地址: MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
Applications@201704
源码地址:
tensorflow版本:wanjinchang/SSH-TensorFlow
相关论文:
RetinaFace: Single-stage Dense Face Localisation in the Wild@20190504
人脸检测-RetinaFace论文解读与应用
mobilnet0.25轻量级检测
Base Module