Anomaly detection models (vision)

[innat]

Short Description

Anomaly detection is an important category of common problem mostly in manufacturing industry. There is a good amount of ML models (and active research) to automate the anomaly detection task. I think it would be good addition to add these models in keras_cv.models.

Papers

• CutPaste by Google.
• FastFlow
• PatchCore
• PaDiM

Benchmark dataset

• MVTec
• BTech

Metrics

• Image and Pixel-level AUC
• Image and Pixel-level F1 score

Existing Implementations

Currently, there are not much strong support available for anomaly detection (vision) in tensorflow.keras. In contrast, a pytorch-lighting based implementations, named anomalib are available, maintained by a OpenVINO team. Currently it offers 10 different anomaly detection (vision) models and actively developing. These model are able to run in supervised (with binary mask) or unsupervised manner (with no mask).

Other Information

If these models are welcomed, we can close this issue and create new ones for these models each.

[bhack]

Some of these models are available in a TF/Keras flavour at https://github.com/Burf/tfdetection

/cc @burf If he Is interested to contribute.

[bhack]

Is MVTec saturated? https://paperswithcode.com/sota/anomaly-detection-on-mvtec-ad

[innat]

Yeah, looks so. Models like patch-core, or fast-flow (transformer based) left less room for improvement on this dataset.

[LukeWood]

At this time we're focusing on ironing out issues in our current offerings before expanding too broadly, lets revisit anomaly detection in the future when the more standard workflows are solved.