https://github.com/gathierry/FastFlow

据说inferencing的时候，各种报错。

https://openvinotoolkit.github.io/anomalib/reference_guide/algorithms/fastflow.html

https://blog.csdn.net/qq_45700830/article/details/122690958

• 传统flow模型，需将2D特征压缩成1D，进行概率估计，损伤了内部的空间关系信息，限制的flow的能力
• 传统flow模型，使用sliding window method，需在大量patch逐个检测异常，复杂度高、推理速度受限
• 本文端到端的推理整个图片，直接输出异常检测和定位结果，推理效率高

1. 2d NF vs 1d NF

   1. DifferNet, image level 1d NF, failed to obtain the exact anomaly localization results since they flattened the outputs of feature extractor
   2. CFLOW-AD: patch levele 1d NF
      1. CFLOW-AD still needs to perform testing phase in the form of a slice window.
   3. FastFlow: image level 2d NF, end-to-end
   4. 这也是很强的position，要求alignment吧？ No
      1. Non-aligned Disturbed MVTec AD Dataset: evaluate our FastFlow (with CaiT) in this new test dataset and we obtain 99.2 image-level AUC and 98.1 pixle-level AUC. There is almost no performance loss compared with the results in original aligned MVTec

   

2. Better than CFLOW-AD

   1. Compared with CFLOW-AD which also uses flow model, our method achieves 1.5× speedup and 2× parameter reduction
   2. alignment problem: CFLOW-AD proposes to use hard code position embedding to leverage the distribution learned by NF, which probably underperforms at more complicated datasets.

3. FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows, 21