[Bug]: Fastdup only accepts one object per image in COCO format

[kasperschnack]

What happened?

When using fastdup with a COCO format dataset, it only recognizes and processes one object per image, first one in the list of annotations, even when multiple objects are present in the annotations.

What did you expect to see?

I expected fastdup to process and recognize all annotated objects per image in the COCO format dataset, allowing for handling images with multiple images.

What version of fastdup were you runnning on?

2.13

What version of Python were you running on?

Python 3.9

Operating System

Mac OS Sonoma 14.7.1 (23H222)

Reproduction steps

1. Prepare a COCO format dataset with multiple objects annotated per image

put all images and annotations.json in imgs/ folder

2. Run fastdup on the prepared dataset using jupyter notebook

import fastdup

fd = fastdup.create(input_dir="imgs/")
fd.run()

fd.explore()

3. Observe the output and visualizations

Notice that only one object per image is recognized and processed

Relevant log output

No errors are shown

Attach a screenshot [Optional]

Only category 133 is shown for image 706. If i switch around the order of the categories such that 54 is first - that is the one that will be registered

Contact Details [Optional]

okps@novonordisk.com

[Tompil3r]

Hi @kasperschnack,
fastdup removes bounding boxes with the exact same coordinates in the same image. Hope this clears things up.

[kasperschnack]

Hi @Tompil3r

Thank you very much for that clarification! Indeed that solves the mystery!

[kasperschnack]

Maybe it would be a good idea to print a warning when this happens? Other people might be confused by this as well

[dbickson]

This is a great idea. We will add this warning and report once completed soon.