← Back to Research

Research & Publications

SCRFD

Sample and Computation Redistribution for Efficient Face Detection

SCRFD is designed for strong face detection accuracy under strict latency and compute budgets across edge, mobile, and server deployments.

Paper Details

SCRFD: Sample and Computation Redistribution for Efficient Face Detection

Publication

ICLR 2022

Authors

Jia Guo, Jiankang Deng, Alexandros Lattas, Stefanos Zafeiriou

Open paper

Research Overview

SCRFD improves efficiency by redistributing both training emphasis and model computation so the detector spends budget where it matters most. The result is a family of detectors that balances accuracy and throughput well for production systems that need real-time or edge-friendly inference.

Production Applications

  • Edge cameras, kiosks, and smart terminals
  • On-device face detection for mobile AI apps
  • High-throughput video preprocessing pipelines
  • Server-side face localization before recognition or liveness analysis

Code Demo

Inspect SCRFD detections with buffalo_l

Load the buffalo_l package, run detection on a crowded image, and print the face boxes and keypoints returned for production pipelines.

demo.py
1import cv2
2from insightface.app import FaceAnalysis
3
4app = FaceAnalysis(name="buffalo_l")
5app.prepare(ctx_id=0, det_size=(640, 640))
6
7img = cv2.imread("crowd.jpg")
8if img is None:
9    raise FileNotFoundError("input image not found")
10
11faces = app.get(img)
12
13for index, face in enumerate(faces):
14    bbox = face.bbox.astype(int).tolist()
15    kps = face.kps.astype(int).tolist()
16    print(f"face {index}: bbox={bbox}")
17    print(f"face {index}: keypoints={kps}")

Key Contributions

Redistributes samples and computation to improve the accuracy-speed tradeoff instead of simply scaling model size upward.

Supports multiple deployment tiers, helping teams choose a detector that matches mobile, embedded, desktop, or server constraints.

Delivers strong WIDER FACE performance while remaining practical for real-time inference pipelines.

Business Contact

Let’s discuss your business needs

Contact InsightFace for model licensing, enterprise deployment support, custom AI development, or strategic partnership discussions.

We only accept inquiries from business email addresses.

Contact via EmailExplore on GitHub