Per-Class Score Threshold

Overview

Different object classes have different score distributions — small dense cells score lower than large isolated ones. A single global threshold under-filters easy classes and over-filters hard ones. Infer the dominant class per image (mode of predicted classes), then apply that class's optimized threshold. Common in cell segmentation where cell types have very different morphologies.

Quick Start

import torch
import numpy as np

THRESHOLDS = [0.15, 0.35, 0.55]  # per class, tuned on validation
MIN_PIXELS = [75, 150, 75]        # per class minimum area

def filter_predictions(predictions):
    scores = predictions['scores']
    classes = predictions['pred_classes']
    masks = predictions['pred_masks']

    # Infer dominant class for this image
    dominant_class = torch.mode(classes)[0].item()

    # Apply class-specific threshold
    keep = scores >= THRESHOLDS[dominant_class]
    return masks[keep], scores[keep], dominant_class

masks, scores, cls = filter_predictions(output['instances'])

Workflow

1. Run inference to get per-instance scores, classes, and masks
2. Compute the mode of predicted classes to determine dominant image class
3. Index into per-class threshold and min-area arrays
4. Filter predictions by the class-specific threshold
5. Optionally apply class-specific min-area filtering post-overlap-resolution

Key Decisions

• Mode vs majority: torch.mode is simple; for mixed-class images, consider per-instance thresholds instead
• Threshold tuning: sweep thresholds per class on validation set, optimize for mAP@IoU
• Homogeneous assumption: works best when images contain mostly one class; for mixed scenes, apply per-instance class thresholds
• Min-area coupling: pair with per-class min_pixels to also filter by class-specific size

References

• Positive score with Detectron 3/3 - Inference