You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[QUESTION] CPDINO : Annotation Strategy, Small Cell Separation, and Occlusion #1506
I'm currently fine-tuning CPDINO for cell instance segmentation and would appreciate some guidance on improving its performance.
1. Annotation completeness and spatial learning
In densely packed cell images, I noticed that many cells in the central regions were missing during inference.
After reviewing my training data, I found that similar dense regions were also incompletely annotated, particularly near the image center.
I added more annotations and retrained from the original pretrained weights (not my previously fine-tuned checkpoint). This noticeably improved detection in those regions.
Could incomplete annotations cause CPDINO to suppress predictions in certain regions, or is this more related to insufficient dense-cell training examples?
2. Explicit labeling of unwanted cells (Outpour)
My images contain circular wells, and I want to exclude cells that overflow outside the well.
Currently, I leave these out-of-well cells unlabeled as background.
Would explicitly labeling them as a separate outpour class and filtering them during inference be more effective?
Does CPDINO support this type of context-dependent multi-class segmentation, or would using a well-boundary mask be more appropriate?
3. Small and densely packed cell separation
I would like to improve the detection and separation of small, tightly packed cells.
Are there recommended training strategies, loss modifications, or inference parameters for improving small-cell recall and reducing merged instances?
4. Occlusion and overlapping cells
I also want to detect partially occluded cells, including cells hidden behind neighboring cells.
Does CPDINO support separating these instances or estimating their occluded boundaries (amodal segmentation)?
If not, are there recommended approaches or architectural modifications to achieve this?
I'm also interested in maintaining robustness across variations in cell morphology, brightness, and texture.
Any guidance on existing features or recommended improvement directions would be greatly appreciated.
Activity