Abnormal floating masks when using SAM 3.1 for MOT, leading to oversized bounding boxes

Hi,

I encountered an issue when using SAM 3.1 in a multi-object tracking (MOT) pipeline.

Specifically, I observed that the generated masks sometimes contain floating or disconnected regions that do not belong to the target object. These artifacts significantly affect the bounding box estimation, making the boxes much larger than the actual object.

This issue appears frequently in challenging scenarios (e.g., occlusion, motion blur, or crowded scenes), and it negatively impacts tracking stability.

Observations:
- Masks occasionally include unrelated regions far from the main object
- The resulting bounding boxes are noticeably oversized
- This leads to degraded tracking performance

Expected behavior:
- Masks should be spatially consistent and tightly aligned with the target object
- No disconnected or floating regions should be included

Questions:
1. Is this a known issue when using SAM for MOT tasks?
2. Are there recommended post-processing steps (e.g., connected component filtering, mask refinement)?
3. Would tuning specific parameters help mitigate this problem?

Any suggestions would be greatly appreciated.

Thanks!

https://github.com/user-attachments/assets/cc8ecd4a-c533-47fa-8dcb-f69b37b456ed

Start by reproducing the oversized boxes in the repository's SAM 3 inference or example notebooks using the attached scenario, focusing on occlusion, motion blur, and crowded scenes. Compare the generated masks with the target objects and document whether the behavior is caused by model output, bounding-box estimation, or post-processing; done means a confirmed cause or a validated mitigation.

Written by the indexing model from the issue text.