Unable to accurately place polygon annotations along the true bottom edge of image

[Tab]

• Project Type: instance segmentation
• Operating System & Browser:
  macOS Tahoe 26.5.2
  Google Chrome Version 149.0.7827.201, Official Build, x86_64
• Project Universe Link or Workspace/Project ID: I’d rather not share
• Do you grant Roboflow Support permission to access your Workspace for troubleshooting? (Yes/No): No
• Roboflow Plan: paid (Basic)

Issue

When annotating assigned images in the Roboflow annotation editor, polygon points/annotations placed along the bottom border of the image do not stay in position/or get placed in the correct position (they shift up a few pixels). If I draw a polygon where part of the vertices sit directly on the x-axis (bottom of the image), the polygon shifts upward by a few pixels when attempting to place a polygon point.

I can sometimes drag the entire polygon downward, but when I do that, the polygon appears to get cut off or constrained near the bottom. After dragging it down, I then have to manually adjust and reshape the polygon because the original shape no longer matches the region I was trying to annotate.

This does not appear to happen with polygons placed against the top border of the image.

my images are standard size images ~1920x1080

Expected Behavior: I should be able to place polygon points directly on the bottom edge of the image, the same way I can place points along other parts of the image.
Actual Behavior: The polygon appears to be constrained or clamped before reaching the true bottom boundary of the image.It behaves as though there is a premature maximum Y-axis limit in the annotation editor, where polygon vertices cannot persist or be placed at the true bottom edge of the image.

Steps to Reproduce

1. Open an assigned image in the Roboflow annotation editor.
2. Select the polygon annotation tool.
3. insert a polygon point at the bottom boundary of the image
4. Observe the polygon point shift upward slightly and no longer aligns with the image border.
5. create a complete polygon anyway
6. Try dragging the entire polygon back down to the bottom border.
7. observe the polygon shifting upward again. and notice that the polygon shape is now cut/cropped a few pixels.
8. attempt to move a polygon point to the bottom of the image, and observe.

For segmentation annotations, it is important that the polygon includes all pixels belonging to the object, and this bug prevents me (anyone else?) from doing that. my ground truth labels are not actually true because of this/annotations aren’t accurately representing objects.

[Automatez]

I’m seeing the same thing - at both the bottom and far right it’s a single pixel width. Based on what I’m reading, if the image is 640 pixels wide, then an annotator can place 640 annotation nodes located on one side of each pixel at 0, 1, 2…638, 639. So you end up “leaving a pixel” at the far edges. Probably minimal impact on the model performance, but if it did end up being critical then A.) you’ll need a killer camera to get that consistent pixel level sensitivity and B.) you could pad the images an extra couple rows of pixels all around to make sure you include it in a consistent way.

Or…we just got punk’d. We’ll see if they say any different!

[Tab]

interesting. Im able to place on the polygon on the right side, so that doesn’t affect me. just the bottom.

Im now wondering if it’s some sort of zoom setting on the browser? I have mine set at 100% but i’m curious what you have yours at?

truthfully I don’t think this will affect the model that much, i’d consider this low priority. it would be an interesting test to see if models performance improves. A lot of my polygons sit on the bottom edge though. so if this gets fixed it would be interesting to see that .01% improvement. lol.

adding padding to the images/remaping the annotations to new image size is a headache I’d rather not have to go through again.

[Automatez]

Browser zoom was at 75%, but I flipped to 100% and still could not access the far right side. I’m surprised you can - I really thought the explanation of the pixel count being the limiter was going to be it.

Agreed - adding padding would be a burden. Even for a new project. Also agree low impact. Maybe someone will kick off an LLM Agent to go test and look for a .01% improvement some day.

[Tab]

today, I was able to replicate the right side pixel cut off. I was attempting to modify a annotation already in the dataset. so I think location of the image within the project doesn’t matter (dataset/ or work-in-progress)

but i think you might be on to something, probably has something to do with how the labels get calculated/normalized for training.

idk, i might be a pixel tyrant here lol.

[Automatez]

Yeah, but what if someone comes along some day and they are like “my entire project was derailed until I found this thread explaining the expected one pixel edge shortage and that revelation saved our company millions!” Just kidding, probably not going to get that.

Really though, I’m the same way - in the long term it’s better to understand than to just ignore.