An implementation of the "Anomaly Detection in 3D Point Clouds using Deep Geometric Descriptors" paper in Python without pre-training with the ModelNet10 dataset and without generating the synthetic data.
After downloading the repo, complete the following:
- Download the MVTec 3D-AD dataset dataset and place its unzipped contents (it should be a folder titled "mvtec_3d_anomaly_detection") in the "datasets" folder.
- Run train script (
py -3 -m train)
If you are having issues or if it is running slowly, run the systemtest script.
Also note that I trained these models on a NVIDIA GeForce RTX 3050 Ti Laptop GPU, so if you make some minor alterations to the train code (number of epochs or fixed_size for example), you can likely train better models than those found here. The models linked below only went through 8 epochs each, and ideally you want 11+ for this kind of model.
Run the test/visualize script with py -3 -m test or py -3 -m test filenamehere.png
Models uploaded to my drive, here: anomaly-detection-3d-models.
