技巧

数据

  • https://zhuanlan.zhihu.com/p/402511359

  • 数据增量: https://github.com/conradry/copy-paste-aug

  • 目标检测增强: 正样本增强:根据类别增强、根据大小增强、增强中加入随机概率、无合适anchor的增强、根据匹配好坏的增强

训练

  • 首先在小backbone如resnet34/efficientnetb0起步,调节loss/augmentation/metric/lr_scheduler

  • 然后再换大模型

分类

检测

mmdetection: https://zhuanlan.zhihu.com/p/337375549

  • EDA:分析图片尺寸信息,标签数量、大小信息(根据数据信息决定:anchor,训练尺度, 感受野) https://github.com/Media-Smart/volkscv/tree/master/volkscv/analyzer/statistics

  • 模型:yolov5,efficientdet,mmdetection(cascade rcnn)

  • 调参:模型,后处理

  • 优化:persudo label,TTA,nms,weighted fusion box

  • https://github.com/felixBrave/ocr_chinese

  • https://github.com/FudanVI/benchmarking-chinese-text-recognition

EDA

  • 图像大小分布

  • gt-box 大小,长宽比分布

  • gt-box与图像大小比例分布

  • 特殊性质

后处理

  • https://www.kaggle.com/shonenkov/wbf-approach-for-ensemble

入门

https://zhuanlan.zhihu.com/p/266864780 https://www.zybuluo.com/huanghaian/note/1743266 https://github.com/LiChenyang-Github/tianchi_Cervical_Cancer_top4 https://www.kaggle.com/code/youhanlee/unet-resnetblock-hypercolumn-deep-supervision-fold

实践

  • https://zhuanlan.zhihu.com/p/409157142

  • https://zhuanlan.zhihu.com/p/363354912

  • https://kulbear.github.io/archives/protein/

  • https://github.com/yl305237731/flexible-yolov5

  • https://github.com/bubbliiiing/yolov5-tf2

MM: https://zhuanlan.zhihu.com/p/337375549?utm_source=zhihu&utm_medium=social&utm_oi=896443966524891136

https://towardsdatascience.com/indian-car-license-plate-detection-using-yolo-v5-ae2574578175

  • https://github.com/facebookresearch/pytorchvideo

Last updated