Sam3 instance bonding

[demoulinv]

This pull request introduces a new "masks bonding" feature to the SAM3 image and video segmentation nodes in Meshroom, allowing overlapping instance masks to be merged using morphological operations. The change is configurable via new parameters and is implemented consistently across all relevant segmentation nodes and processing pipelines.

Feature: Masks Bonding for Overlapping Instances

• Added new parameters enableBonding (bool) and bondingKernelSize (int) to the following nodes: ImageSegmentationSam3, VideoSegmentationSam3, VideoSegmentationSam3Text, and VideoSegmentationSam3Boxes, enabling users to control bonding behavior and kernel size for morphological processing.

Implementation: Mask Bonding Logic

• Introduced a new utility function bond_masks in sam3Utils.py that performs dilation, conflict region detection, and morphological closing to merge overlapping masks, controlled by kernel size parameters.

Integration: Processing Pipeline Updates

• Updated mask processing logic in all relevant processChunk methods to use the new bonding feature: when enabled and multiple masks are present, masks are merged using the bonding kernel size. This affects both forward and backward propagation in video segmentation, as well as tiling and text-prompted segmentation.

Documentation and Versioning

• Updated documentation strings and comments to reflect the new bonding feature, and incremented the version of VideoSegmentationSam3Boxes to 4.1.

These changes collectively provide more robust control over how overlapping segmentation masks are handled, improving the quality of mask outputs in scenarios with touching or overlapping objects.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.