BANOS

Set of metrics to assess behavior annotations of videos.

BANOS (Behavior ANnOtation Score) evaluates algorithmic behavior annotations against a ground truth (typically human annotations). Designed for ethology and computer vision workflows.

Traditional frame-based metrics miss important temporal qualities. BANOS addresses this with four complementary metrics:

1. Detection Accuracy (DA) — Precision, Recall, F1 at the bout level

2. Segment Overlap (SO) — Temporal Intersection over Union (tIoU)

3. Temporal Precision (TP) — 1 / (1 + |start deviation| + |end deviation|)

4. Intra-bout Continuity (IC) — 1 - (label switches / bout length)

Usage:

run('setup.m'); % add to path

metrics = BANOS_score(pred, gt); % single recording

[g, o] = BANOS_score(dataStruct); % multiple recordings

Inputs are cell arrays: row 1 = behavior headers, rows 2+ = binary (0/1) frame data.

A Python implementation (pip install banos) with identical results is available on PyPI. See GitHub for demos and tutorials.