Regenerating mvtec_loco_composition_maps are completely different

[HU13-SVG]

Why do the images generated when regenerating the mvtec_loco_composition_maps differ completely from the images provided for download?

[HU13-SVG]

Could it be sam_vit_h_4b8939.pth? The code says sam_hq_vit_h.pth, but the download is sam_vit_h_4b8939.pth. Is there any difference?

[MaticFuc]

I will recheck sometime this week to see if I accidentally added a bug in the release code. Could you show what kinds of pseudo-labels you achieve?

[HU13-SVG]

Below are images generate from python create_pseudo_labels.py --category screw_bag，no code changes. I found that there is a little difference from what was reported in the paper.
Furthermore, a tiny mistake, when run python train_composition_segmentation_model.py --category screw_bag, there is no images saved in ./mvtec_loco_composition_maps/screw_bag/train/good/.

[MaticFuc]

I think I found the bug. I didn't have the time to rerun the experiment, but it seems I accidentally set the wrong pseudo-label path in train_composition_model.py. I have fixed this and uploaded it to GitHub.

I will check if there is anything extra wrong in the next few days.

As the pseudo-label creation involves sampling data, the final composition maps will change slightly. More or less, they will stay the same.

[HU13-SVG]

1. In train_composition_segmentation_model.py, line 171, default=".data/mvtec_loco_noisy_composition_maps", add a / to be "./data/mvtec_loco_noisy_composition_maps"

2. Furthermore, check line 135 "predicted_mask = model(img).softmax(dim=1)" if it's right. Because nn.CrossEntropyLoss() already contains the calculation logic of the Softmax operation.

[zhiqing0205]

same question

[HU13-SVG]

I guess the author set up code protection intentionally. I have read DRAEM of the author team and the code level is tested. The author may be preparing for CVPR 2026.

[MaticFuc]

The code for generating composition maps is working as intended now. There was a small bug in the argparser. The default path was set to .data/... instead of ./data/.... Due to that, the model you are using has random weights and was not trained on any images. Hence, the images you are showing.

For better results, you could also try DINOv3 or some more advanced models for pseudo-label creation.

@HU13-SVG, no code protection was set intentionally. If such a thing were intended, I would not release the code. This codebase was merged from three different ones, which led to new bugs during the merging. I have unfortunately missed them.

But I am actually preparing something for CVPR 2026 (unrelated to Logical Anomaly Detection), so I did not have the time to check the code for SALAD in such detail as I would've wished.

[zhiqing0205]

The author’s latest commit has improved things, but there are still minor inconsistencies compared with the version available for download.

[oylz]

Could it be sam_vit_h_4b8939.pth? The code says sam_hq_vit_h.pth, but the download is sam_vit_h_4b8939.pth. Is there any difference?

sam_hq_vit_h.pth is from here:sam-hq:Segment Anything in High Quality [NeurIPS 2023]

arxiv.org/abs/2306.01567