README.md

metadata

title: Face Detection (SCRFD comparison)
emoji: 🙂
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false

Face Detection — SCRFD model comparison

Upload one image; it is run through all three SCRFD face detectors at once and the annotated results are shown side by side so you can compare them in a single pass — no model picker.

Model	File	Input	Notes
SCRFD-500MF	models/det_500m.onnx	640×640	lighter / faster
SCRFD-2.5GF	models/det_2.5g.onnx	640×640	heavier / usually higher recall
SCRFD @ 480	models/det_480.onnx	480×480	same SCRFD-500MF weights as det_500m, exported at a smaller input — fastest, but drops more small/distant faces

The third panel isolates the effect of input resolution: det_480 uses the very same SCRFD-500MF network as det_500m, just fed a 480×480 letterbox instead of 640×640, so any difference you see is purely down to input size (speed vs. reach on small faces).

Each panel draws every detected face (bounding box + confidence) and its title bar reports the face count and inference time; the 5-point landmark coordinates for each face are listed as text below the panel. CPU-only ONNX Runtime — runs on a free Hugging Face Space.

How it works

1. The image is letterboxed to each model's input size (640×640 for det_500m / det_2.5g, 480×480 for det_480), aspect ratio preserved.
2. All three SCRFD models run in ONNX Runtime (CPU).
3. scrfd.py decodes the anchor-based outputs (3 strides × {score, bbox, kps}), applies IoU-NMS, and maps boxes back to the original image.

scrfd.py is a self-contained re-implementation of InsightFace's SCRFD post-processing, so the raw .onnx files load directly — no insightface package or model-pack directory needed.

Run locally

pip install -r requirements.txt
python app.py

Then open the URL shown in the terminal (e.g. http://127.0.0.1:7860).

Models

det_500m.onnx and det_2.5g.onnx are the standard InsightFace SCRFD detection checkpoints (both with 5-point landmarks). det_480.onnx is the same SCRFD-500MF network re-exported with a fixed 480×480 input. They live in models/.