Label information corresponding to the color of segment in the rgb mask

[sanngu5]

Thank you for providing the amazing dataset.

In the rgb mask of the dataset, is there any label information for each color?

I'm curious about this beacause the semantically same defect in DS MVTec and Synthetic data of DS MVTec is shown in different colors.

[zhifeichen097]

Hi, we appreciate your interest in our dataset, the label information can be found in https://huggingface.co/datasets/DefectSpectrum/Defect_Spectrum/blob/main/DS-MVTec/DS-MVTec.md; as for the different color, please use the convert_mask in the synthetic dataset. The rgb_mask in synthetic dataset is generated by assigning a random color, the true label can be found in convert_mask folder(https://huggingface.co/datasets/DefectSpectrum/Defect_Spectrum/tree/main/synthetic_MVTec/bottle/masks_converted_dir). Let me know if you have further questions.

[sanngu5]

Thanks for the answer!
I checked the label information on DS-MVTec.md that you pointed out, but I have one more question.

‘cable’: {“bent_wire”: 1, “cable_swap”: 2, “combined”: 3, “cut_inner_insulation”: 4, “cut_outer_insulation”: 5, “missing_cable”: 6, “missing_wire”: 7, “poke_insulation”: 8},

It appears that the combined class has been deleted for the pill class, whereas how can combined be defined as a single defect class for the cable class.

Also, some data in the original dataset seems to have been removed in the DS version (e.g. flip class of metal_nut)

I would really appreciate it if you could answer these two questions!

[AndysonYs]

Hi, thanks for the follow-up!

I checked the cable data again and noticed that label “3” is actually also referring to “cut_inner_insulation”—there is no defect class named “combined.” In our categorization, we assigned a separate label to each distinct defect type. For convenience during annotation, you can temporarily used this mapping:

‘cable’: {“bent_wire”: 1, “cable_swap”: 2, “cut_inner_insulation”: 3, “cut_inner_insulation”: 4, “cut_outer_insulation”: 5, “missing_cable”: 6, “missing_wire”: 7, “poke_insulation”: 8}

As for the “flip” defect in metal_nut, we removed it because it’s more of a defect classification issue rather than a segmentation one. The entire object would be marked as the defect mask, which doesn’t align well with the segmentation-focused scope of our work.

Hope this clears things up! Let me know if you have more questions.

[sanngu5]

Thank you for your response.

1. When I reviewed the mask of the cable class in DS-MVTec, it seems that class 4 represents the cut_outer area, not the cut_inner_insulation, so I think this mapping is correct, but is it?

"cable": {'bent_wire': 1, 'cable_swap': 2, 'cut_inner_insulation': 3, 'cut_outer_insulation': 4, 'missing_cable': 5, 'missing_wire': 6, 'poke_insulation': 7}

2. Something similar issue happened with the wood class in DS-MVTec. In this case, The dataset has 'combined' but the mapping doesn't have 'combined'. (but has 'crack', which is not in dataset)