How to *not interpolate* the input image for labeling

[notAstereo]

Hello,
I'd like to know if there is a way to tell labelme to not interpolate the image for labeling.
I want to see the individual pixels when I add my bounding boxes.

[gaspardringuenet]

Same issue here. Has someone found a way since?

[gaspardringuenet]

My use case might be a bit peculiar as I am not working on natural images. I use LabelMe to annotate scientific data (here oceanographic sonar echogram i.e. spatio-temporal images showing the backscattering intensity). As illustrated in the second screenshot, strong intensity variations can occur at the pixel level, and they are biologically meaningful. Annotation can often be done despite the filter. However:

• In some case blurring prevents the user from seing the start / end of an object (here typically fish / other organisms assembled in a patch or layer)
• A gaussian filtering for this kind of data visualization is considered wrong as it alters real measurements

A example of an echogram image annotated using LabelMe. The x axis is ping time, the y axis is depth, and the intensity is mapped to the volume backscattering variable (a measure of the presence of organisms in the sampled volume) Zoom an a specific feature with high pixel-wise intensity variations

[wkentaro]

interesting. can you please share the original image so that i can test myself (and address and verify the remedy)?

[gaspardringuenet]

Here it is !

And here's a zoom on the region of interest I used as an example. It remains somehow blurry but this time it is linked to the underlying data.

[wkentaro]

looks i reproduced. will dig deeper soon

[wkentaro]

This was generated by AI during triage.

Tracked as #2142 (labeled type: feature + ready-for-human). The blur comes from the canvas always painting the image with the SmoothPixmapTransform render hint, so the image is bilinearly interpolated past 100% zoom. The fix is to add an opt-in nearest-neighbor mode for the image layer (shape outlines stay antialiased). @gaspardringuenet you can follow #2142 for progress.