特征匹配 https://zhuanlan.zhihu.com/p/52140541 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108078#latest-621878
ensemble技巧 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107716#latest-624046 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/111457#latest-642578
这个链接提到训练时长的问题,或许需要保存中间结果 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108554#latest-626181
提到了Dice-Score https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/101465#latest-586178
一篇检测锈斑的论文 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/101471#latest-625980 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109297#latest-631198 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108821#latest-629610 https://software.intel.com/en-us/articles/use-machine-learning-to-detect-defects-on-the-steel-surface
引导性链接 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/101969#latest-641353 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/103296#latest-640460
关注图像角落里的第一个像素的坐标到底是(1,1)还是(0,1) https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/102146#latest-589715
提到了一篇论文讨论了语义分割里面的不同类型的loss https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/102386#latest-625072 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/110536#latest-639400 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108206#latest-635042
提供了一些网络 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/105296#latest-606287
下面这几个没有完全看懂 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/103861#latest-600125 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/103367#latest-639821 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106477#latest-642453 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109423#latest-630712 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108270#latest-629664 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107889#latest-631449
半监督 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/110426#latest-641084
提到了数据增强 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/104850#latest-606137 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109227#latest-640539
貌似是使用了条件随机场 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106086#latest-613534
蛙哥说先判断一个像素是不是锈斑,然后判断是第几类 然后提到不要使用所有数据,那样反而会让得分低下 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106099#latest-629814
照片一致,但是标签不一致 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107053#latest-621775
pool大小的调整建议 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106952#latest-620343
新手包 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106462#latest-641632
说法是34层的resnet最好 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108949#latest-636914
以前的语义分割冠军方案 https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108308#latest-625068
椒盐噪声和对抗验证 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/111119#latest-640192 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106834#latest-633503 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108790#latest-627471
找到很多子类 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/110363#latest-638823
提出一个问题: 使用预训练的网络,但是预训练的图片和当前的图片不一样的时候如何处理?(帖子内容我没看,其实就是修改最后一层) https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107246#latest-618321
kaggle在语义分割中的得分机制dice-score https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/110188#latest-642222
貌似需要扔掉一些图片 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109673#latest-637866
一大堆神经网络的论文 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109370#latest-631305
提到了IOU https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109847#latest-632505
语义分割网络回顾 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109318#latest-629292
下面这个似乎非常重要，据说只要移除False Positive，就可以获得0.9117 https://www.kaggle.com/evgenyshtepin/severstal-mlcomp-catalyst-infer-0-90726 https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/106462#latest-634450
这个EDA做的很漂亮 https://www.kaggle.com/avirald/clear-mask-visualization-and-simple-eda
这个链接提到IoU是一种 loss https://www.kaggle.com/rishabhiitbhu/unet-starter-kernel-pytorch-lb-0-88
