Dear authors,
I really appreciate your great work on SAM‑3. I have attempted to reproduce the reported semantic‑segmentation mIoU results in the original paper but cannot match the paper’s numbers after multiple trials. Could you kindly clarify the following implementation details? What is the exact computation pipeline for semantic‑segmentation mIoU described in the paper?
Are semantic segmentation predictions generated from the semantic‑segmentation head, or assembled from instance‑segmentation‑head outputs?
What mask‑filtering strategy was used during semantic‑segmentation evaluation?
How are IoU values averaged across different categories or different samples?
Dear authors,
I really appreciate your great work on SAM‑3. I have attempted to reproduce the reported semantic‑segmentation mIoU results in the original paper but cannot match the paper’s numbers after multiple trials. Could you kindly clarify the following implementation details? What is the exact computation pipeline for semantic‑segmentation mIoU described in the paper?
Are semantic segmentation predictions generated from the semantic‑segmentation head, or assembled from instance‑segmentation‑head outputs?
What mask‑filtering strategy was used during semantic‑segmentation evaluation?
How are IoU values averaged across different categories or different samples?