Throughput benchmarks

Contents

Throughput benchmarks#

Tracks per-release wall-time of Detector and MPDetector over time. Each entry is a single run produced by python scripts/bench_detectors.py --markdown. Accuracy benchmarks live in accuracy.md.

Methodology#

scripts/bench_detectors.py measures the detection + landmark + AU + emotion + identity path on reproducible test fixtures from feat/tests/data/:

  • single_face.mp4 (72 frames, 1 face/frame)

  • WolfgangLanger_Pexels.mp4 (472 frames, 1 face/frame)

  • multi_face.jpg × 16 = 80 faces

Three configurations are timed head-to-head:

  1. Detector(face_model='img2pose', au_model='xgb', emotion_model='resmasknet', identity_model='arcface')

  2. Detector(face_model='retinaface', au_model='xgb', emotion_model='resmasknet', identity_model='arcface')

  3. MPDetector(face_model='retinaface', landmark_model='mp_facemesh_v2', au_model='mp_blendshapes', emotion_model='resmasknet', identity_model='arcface')

Swept axes: device × batch_size × num_workers.

History#

date

run

2026-05-14

2026-05-14-c716340.md

2026-05-04

2026-05-04-f44ccb1.md

2026-05-03

2026-05-03-d71c0d7.md

2026-05-03

2026-05-03-864962c.md

2026-05-03

2026-05-03-437b651.md

2026-05-03

2026-05-03-09980f9.md
Contents