RetinaFace
Single-shot face detector that also predicts five landmarks and dense 3D face vertices — the detection front-end behind countless recognition pipelines.
About
RetinaFace is a one-stage face detector that treats face localization as multi-task learning: every anchor predicts a face score, a bounding box, five facial landmarks, and pixel-wise dense 3D face vertices. The landmark branch was the key trick — manually annotating five landmarks on WIDER FACE gave the detector extra supervision that paid off hardest on tiny, difficult faces.
It became the default detection stage for ArcFace/InsightFace pipelines because one model gives you detection, alignment-ready landmarks, and 3D shape at once. The official weights live in the InsightFace repository, and third-party reimplementations exist in TensorFlow, PyTorch, and ONNX, all reproducing the paper's WIDER FACE numbers within a point or two.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: WIDER FACE validation AP 96.9% easy / 96.1% medium / 91.8% hard (test set: 96.3 / 95.6 / 91.4), setting the hard-set record at publication. Third-party reimplementations (e.g. TensorFlow 2 port) report 95.6 / 94.6 / 88.5, close to the original — one of the few face detectors with genuine community reproduction.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified, metrics-v1). Evidence: peer-reviewed CVPR 2020 paper (arXiv 1905.00641) by the InsightFace team; official weights in the InsightFace repo; multiple community reimplementations reproduce the numbers within a couple of points. Strengths: strong provenance, reproduced accuracy, enormous downstream adoption as a pipeline front-end. Docked for: (1) model weights are research releases with no versioned support; (2) officially superseded by the team's own SCRFD for efficient detection. License: InsightFace code is MIT; weight licensing is research-oriented.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · CVPR 2020 (arXiv 1905.00641) by the InsightFace team; single-shot detector with box + 5 landmarks + dense 3D vertices.
- · WIDER FACE validation AP 96.9/96.1/91.8 (easy/med/hard) provider-reported; community TF2 port reproduces within ~1-2 pts.
- · Superseded by the team's own SCRFD for efficient detection.