Stitching of segmentation volumes

[juampabloheras]

Hello,
Thank you for providing your implementation, we are very excited to try out your method.

We have run the segmentation code and realized that it produces small sub-volume segmentations which need to be patched into the large volume. Would it be possible to share the script you used to merge all sub-volumes into the final segmentation?

Thanks,
Juampablo

[JINcloud6]

This issue has been resolved, and the corresponding pull request has been merged.

What was added

• Large 2D image segmentation via tiled inference in seg_large_image

  • Uses 512 × 512 tiles with a 32-pixel context overlap.
  • Only the non-overlapping center regions are merged into the final mask to reduce boundary artifacts.

• 3D volume deflickering

  • Normalizes each slice against the middle slice using the 1st and 99th intensity percentiles.
  • Preserves the original input dtype.

• Dense 3D TriSAM segmentation

  • trisam_3d_segment_dense now performs tracking along all three orthogonal axes with:

    track_axis=(0, 1, 2)

  • The resulting masks are merged into a shared 3D segmentation volume.

• Missing prompt utilities

  • get_next_point_prompt
  • is_border
  • move_prompt

Running TriSAM

Install the required dependencies:

pip install numpy h5py torch scipy scikit-image opencv-python cc3d pytest
pip install git+https://github.com/facebookresearch/segment-anything.git

Configure the paths:

export TRISAM_DATA_ROOT=/path/to/data
export TRISAM_SAM_CHECKPOINT=/path/to/checkpoints/sam_vit_h_4b8939.pth
export TRISAM_OUTPUT_DIR=/path/to/output

Then run, for example:

python model/main.py macaque 0.98 0.8 "point+bbox" 0

Large-image preprocessing can also be used independently through seg_large_image.

Tests

Run:

python -m pytest -q

The added tests cover tiled reconstruction, deflickering, input validation and preservation, prompt utilities, border detection, and dense three-axis tracking.

Thanks for reporting the issue. Please feel free to reopen it or create a follow-up issue if you encounter any remaining edge cases.