Hi, thank you for open-sourcing this valuable project and dataset.
While checking several samples in STRIDE-QA, I noticed what seems to be a possible region ID alignment issue. In some cases, the region IDs referenced in the QA annotations do not appear to match the corresponding 2D bounding boxes or segmentation masks at the instance level.
More specifically, the problem is not only visible from fine-grained inspection, but can also be inferred from very prominent spatial cues. For example, in one sample, the description assigned to id5 seems to correspond much more naturally to the target labeled as id4 in the image. This suggests that there may be an instance-level indexing mismatch somewhere in the annotation pipeline.
This issue could affect tasks that rely on precise grounding between language annotations and visual regions, especially for spatial reasoning and region-aware evaluation.
Would it be possible to check whether there is:
an offset or permutation in the region ID assignment,
a mismatch between QA annotations and the exported box/mask annotations, or
a bug introduced during preprocessing or annotation conversion?
I am raising this here in case it is helpful for improving the dataset. Thank you again for your great work.
Hi, thank you for open-sourcing this valuable project and dataset.
While checking several samples in STRIDE-QA, I noticed what seems to be a possible region ID alignment issue. In some cases, the region IDs referenced in the QA annotations do not appear to match the corresponding 2D bounding boxes or segmentation masks at the instance level.
More specifically, the problem is not only visible from fine-grained inspection, but can also be inferred from very prominent spatial cues. For example, in one sample, the description assigned to id5 seems to correspond much more naturally to the target labeled as id4 in the image. This suggests that there may be an instance-level indexing mismatch somewhere in the annotation pipeline.
This issue could affect tasks that rely on precise grounding between language annotations and visual regions, especially for spatial reasoning and region-aware evaluation.
Would it be possible to check whether there is:
an offset or permutation in the region ID assignment,
a mismatch between QA annotations and the exported box/mask annotations, or
a bug introduced during preprocessing or annotation conversion?
I am raising this here in case it is helpful for improving the dataset. Thank you again for your great work.