Multi-Annotator Bbox Consensus

Overview

Datasets annotated by multiple experts (radiologists, pathologists) contain duplicate and conflicting bounding boxes for the same object. Using all raw annotations introduces noise; using only one annotator discards information. This technique groups overlapping same-class boxes by IoU, then computes a consensus box — either the intersection (inner box, conservative) or union (outer box, inclusive). Reduces annotation noise and improves detector training, especially when annotators have varying skill levels.

Quick Start

import numpy as np

def compute_iou(box1, box2):
    x1 = max(box1[0], box2[0])
    y1 = max(box1[1], box2[1])
    x2 = min(box1[2], box2[2])
    y2 = min(box1[3], box2[3])
    if x2 < x1 or y2 < y1:
        return 0.0
    inter = (x2 - x1) * (y2 - y1)
    area1 = (box1[2]-box1[0]) * (box1[3]-box1[1])
    area2 = (box2[2]-box2[0]) * (box2[3]-box2[1])
    return inter / (area1 + area2 - inter)

def consensus_box(boxes, mode='inner'):
    """Compute consensus from overlapping boxes."""
    boxes = np.array(boxes)
    if mode == 'inner':  # intersection
        return [boxes[:,0].max(), boxes[:,1].max(),
                boxes[:,2].min(), boxes[:,3].min()]
    else:  # union
        return [boxes[:,0].min(), boxes[:,1].min(),
                boxes[:,2].max(), boxes[:,3].max()]

def merge_annotations(annots, iou_thresh=0.0, mode='inner'):
    """Merge same-class overlapping boxes from multiple annotators."""
    merged = []
    used = set()
    for i, (cls_i, box_i) in enumerate(annots):
        if i in used:
            continue
        group = [box_i]
        for j, (cls_j, box_j) in enumerate(annots):
            if j <= i or j in used or cls_i != cls_j:
                continue
            if compute_iou(box_i, box_j) > iou_thresh:
                group.append(box_j)
                used.add(j)
        merged.append((cls_i, consensus_box(group, mode)))
    return merged

Workflow

1. Group annotations by image_id
2. For each same-class pair, compute IoU
3. Cluster overlapping boxes (IoU > threshold)
4. Compute consensus box per cluster (inner or outer)
5. Use consensus annotations for training

Key Decisions

• Inner vs outer: Inner (intersection) is conservative, reduces box size; outer (union) is inclusive
• IoU threshold: 0.0 merges any overlapping boxes; 0.3–0.5 requires significant overlap
• Minimum annotators: Optionally require 2+ annotators to agree before keeping a box
• Weighted average: Use annotator-weighted mean of coordinates instead of hard intersection

References

• Visual In-Depth EDA – VinBigData