Information about tiling on custom dataset

[LorenzoF6]

Hi, i'm new in the anomaly detection tasks and i'm trying to understand the mechanism behind tiling on the anomalib's library and i have several question to ask about it.

• all model support tiling or some of them ? which one ?
• if i have a custom dataset of 1024x2014 image resolution, tiling is needed to improve the model's performance?
  thank you all for the response in advance

[blaz-r]

Check the docs for tiling info.

But to answer your questions:

• models currently supported are: Padim, Patchcore, STFPM, Reverse Distillation.
• main goal of tiling is to enable processing of higher resolutions - this usually results in a better performance, however if you compare against the same resolution that is not tiled, then usually this is not the case.

[blaz-r]

Hi @onkarkris, I'll have to look into this, but I suspect that the problem might be in center crop. I am now working on refactoring this code to work with v1, I'll also make sure it works there.

For the time being, can you try to use the input size of 1024 without center-crop.

[blaz-r]

I also noticed that you haven't shared the ens_config.yaml file that you are using. Can you also send that please?

[onkarkris]

@blaz-r thanks it would be great if you can find the issue! I will check and let you know if turning off crop solve the issue

ens_config -
`tiling:
tile_size: 256
stride: 256

predictions:
storage: memory # how the predictions are stored, options: [memory, file_system, memory_downscaled]
downscale_factor: 0.5 # factor by which data is downscaled in case of "memory_downscaled" storage

post_processing:
normalization: image # on what level we normalize, options: [tile, image, none]
smooth_joins:
apply: True # if this is applied, area around tile joins are is smoothed
sigma: 2 # sigma of gaussian filter used to smooth this area
width: 0.1 # width factor, multiplied by tile dimension gives the region width around join which will be smoothed

overrides metrics from model config

metrics:
image:
- F1Score
- AUROC
pixel:
- F1Score
- AUROC
threshold:
stage: image # stage at which we apply threshold, options: [tile, image]
method: adaptive # options: [adaptive, manual]
manual_image: null
manual_pixel: null

overrides visualization from model config

visualization:
show_images: False # show images on the screen
save_images: True # save images to the file system
image_save_path: null # path to which images will be saved
mode: full # options: ["full", "simple"]
`

[blaz-r]

Hi, I tried running with 1024 and I was able to train. I'm pretty sure the problem is in center crop. I'll fix that for the v1 refactor, for now I recommend you don't use that.

[blaz-r]

I also managed to reproduce the issue, it is center-crop related, as the tiler takes image_size as input, but it should take center-crop if that is set. As I said, for the time being, avoid center-crop and just set the image size to the desired value - 1024 in your case.