Appendix A — How duplicate masks are removed

SAM2 automatic segmentation can sometimes return multiple masks for the same particle. This can happen when the same object is detected from different prompt points or crop regions. If those masks are counted as separate particles, the final particle count becomes too high.

When deduplicate masks is enabled, this app compares every pair of SAM2 masks and removes likely duplicates before creating the editable label layer.

Overlap measures

For two masks A and B, the app first calculates the number of pixels in their intersection, their individual areas, and their union. It then uses two overlap measures.

Containment measures how much the smaller mask is included in the other mask:

contain = intersection(A, B) / min(area(A), area(B))

This is useful when one mask is mostly inside another mask. For example, if SAM2 outputs both a full particle mask and a smaller mask covering the center of the same particle, containment becomes high even when IoU is not very high.

IoU (Intersection over Union) measures how similar the two masks are as whole regions:

iou = intersection(A, B) / union(A, B)

This is useful when two masks have similar size and shape.

Duplicate decision

A pair of masks is treated as a duplicate candidate when either of the following is true:

contain >= contain thresh
iou >= iou thresh

The default settings are:

contain thresh: 0.90
iou thresh: 0.70

Using both measures helps cover two common duplicate patterns:

One mask is almost contained in another mask
Two masks cover nearly the same region

Which mask is kept

When two masks are judged to be duplicates, the app compares the score specified by score key. By default, this is predicted_iou, which is SAM2’s predicted mask quality score. The mask with the higher score is kept, and the lower-scoring mask is removed.

If the scores are equal, the earlier mask in the sorted mask list is kept. Masks are sorted by centroid position before deduplication, so label IDs remain relatively stable from top-left to bottom-right.

Parameter adjustment guide

If too many masks are removed, increase contain thresh or iou thresh. This makes duplicate removal more conservative.

If obvious duplicate masks remain, decrease contain thresh or iou thresh. This makes duplicate removal more aggressive.

For particle-counting workflows, it is usually better to start with the default values, inspect the overlay, and adjust only when duplicate detections or over-removal are visible.

--- title: "How duplicate masks are removed" --- SAM2 automatic segmentation can sometimes return multiple masks for the same particle. This can happen when the same object is detected from different prompt points or crop regions. If those masks are counted as separate particles, the final particle count becomes too high. When `deduplicate masks` is enabled, this app compares every pair of SAM2 masks and removes likely duplicates before creating the editable label layer. ## Overlap measures For two masks A and B, the app first calculates the number of pixels in their intersection, their individual areas, and their union. It then uses two overlap measures. **Containment** measures how much the smaller mask is included in the other mask: ```text contain = intersection(A, B) / min(area(A), area(B)) ``` This is useful when one mask is mostly inside another mask. For example, if SAM2 outputs both a full particle mask and a smaller mask covering the center of the same particle, containment becomes high even when IoU is not very high. **IoU (Intersection over Union)** measures how similar the two masks are as whole regions: ```text iou = intersection(A, B) / union(A, B) ``` This is useful when two masks have similar size and shape. ## Duplicate decision A pair of masks is treated as a duplicate candidate when either of the following is true: ```text contain >= contain thresh iou >= iou thresh ``` The default settings are: - `contain thresh`: **0.90** - `iou thresh`: **0.70** Using both measures helps cover two common duplicate patterns: - One mask is almost contained in another mask - Two masks cover nearly the same region ## Which mask is kept When two masks are judged to be duplicates, the app compares the score specified by `score key`. By default, this is `predicted_iou`, which is SAM2's predicted mask quality score. The mask with the higher score is kept, and the lower-scoring mask is removed. If the scores are equal, the earlier mask in the sorted mask list is kept. Masks are sorted by centroid position before deduplication, so label IDs remain relatively stable from top-left to bottom-right. ## Parameter adjustment guide If too many masks are removed, increase `contain thresh` or `iou thresh`. This makes duplicate removal more conservative. If obvious duplicate masks remain, decrease `contain thresh` or `iou thresh`. This makes duplicate removal more aggressive. For particle-counting workflows, it is usually better to start with the default values, inspect the overlay, and adjust only when duplicate detections or over-removal are visible.