---
license: mit
license_link: LICENSE
library_name: openvino
pipeline_tag: image-classification
tags:
  - openvino
  - intel
  - face-detection
  - face-reidentification
  - edge-ai
  - metro
  - dlstreamer
language:
  - en
---

# Facial Recognition

| Property | Value |
|---|---|
| **Category** | Face Detection + Re-Identification |
| **Base Model** | [face-detection-adas-0001](https://docs.openvino.ai/2024/omz_models_model_face_detection_adas_0001.html) + [face-reidentification-retail-0095](https://docs.openvino.ai/2024/omz_models_model_face_reidentification_retail_0095.html) (Open Model Zoo) |
| **Source Framework** | Caffe / PyTorch (Open Model Zoo) |
| **Supported Precisions** | FP32, FP16 |
| **Inference Engine** | OpenVINO |
| **Hardware** | CPU, GPU, NPU |
| **Detected Class(es)** | Human faces (detection) + 256-d face embeddings (re-identification) |

---

## Overview

Facial Recognition is a Metro Analytics use case that detects human faces in
images and video and computes a 256-dimensional embedding vector for each face,
enabling enrollment, search, and identification against a known gallery.

The pipeline composes two Intel Open Model Zoo models:

- **face-detection-adas-0001** -- an SSD-based face detector optimized for
  automotive and surveillance cameras (FP16, 384x672 input).
- **face-reidentification-retail-0095** -- a compact CNN that maps a cropped
  face to a 256-d embedding; cosine similarity between embeddings determines
  identity.

These models are well-tested with OpenVINO Runtime and Intel DLStreamer's
`gvadetect` + `gvaclassify` pipeline.

Typical Metro deployments include:

- **Access Control** -- match employees or authorized personnel against an enrollment gallery.
- **VIP Identification** -- recognize known individuals in a crowd.
- **Search and Forensics** -- find a person of interest across multiple camera feeds.
- **Attendance Tracking** -- log when enrolled individuals enter or leave a facility.

> **Privacy Note:** Facial recognition involves biometric data.
> Ensure your deployment complies with applicable privacy regulations
> (GDPR, BIPA, etc.) and has proper consent mechanisms in place.

---

## Prerequisites

- Python 3.11+
- [Install OpenVINO](https://docs.openvino.ai/2026/get-started/install-openvino.html) (latest version)
- [Install Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/get_started/install/install_guide_ubuntu.html) (latest version)

Create and activate a Python virtual environment before running the scripts:

```bash
python3 -m venv .venv --system-site-packages
source .venv/bin/activate
```

> **Note:** The `--system-site-packages` flag is required so the virtual
> environment can access the system-installed OpenVINO and DLStreamer Python
> packages.

---

## Getting Started

### Download Models

Run the provided script to download the face detection and re-identification
models from the Open Model Zoo:

```bash
chmod +x export_and_quantize.sh
./export_and_quantize.sh
```

The script performs the following steps:

1. Installs `openvino`.
2. Downloads `face-detection-adas-0001` (FP16) into `./intel/face-detection-adas-0001/FP16/`.
3. Downloads `face-reidentification-retail-0095` (FP16) into `./intel/face-reidentification-retail-0095/FP16/`.
4. Downloads a sample test video (`test_video.mp4`).

### OpenVINO Sample

The sample below runs recognition on the sample video. It detects every face,
computes a 256-d embedding, and matches it against a gallery of previously seen
people. Each new person is enrolled and assigned a numeric ID; when the same
person is seen again, the gallery returns their existing ID. Every face is
annotated with its `ID <n>`, and the result is saved to `output_openvino.mp4`.
Change the `device` string to run on CPU, GPU, or NPU.

```python
import cv2
import numpy as np
import openvino as ov

DETECTION_MODEL = "intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml"
REID_MODEL = "intel/face-reidentification-retail-0095/FP16/face-reidentification-retail-0095.xml"
INPUT_VIDEO = "test_video.mp4"
CONF_THRESHOLD = 0.6
MATCH_THRESHOLD = 0.5

core = ov.Core()

# Change device to "GPU" or "NPU" to run on integrated GPU or NPU.
det_model = core.compile_model(core.read_model(DETECTION_MODEL), "CPU")
reid_model = core.compile_model(core.read_model(REID_MODEL), "CPU")

det_input = det_model.input(0)
det_h, det_w = det_input.shape[2], det_input.shape[3]
reid_input = reid_model.input(0)
reid_h, reid_w = reid_input.shape[2], reid_input.shape[3]


def detect_faces(img):
    h0, w0 = img.shape[:2]
    blob = cv2.resize(img, (det_w, det_h))
    blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
    detections = det_model([blob])[det_model.output(0)][0][0]
    boxes = []
    for det in detections:
        if float(det[2]) < CONF_THRESHOLD:
            continue
        x1 = max(0, int(det[3] * w0))
        y1 = max(0, int(det[4] * h0))
        x2 = min(w0, int(det[5] * w0))
        y2 = min(h0, int(det[6] * h0))
        if x2 > x1 and y2 > y1:
            boxes.append((x1, y1, x2, y2))
    return boxes


def get_embedding(img, bbox):
    x1, y1, x2, y2 = bbox
    crop = img[y1:y2, x1:x2]
    blob = cv2.resize(crop, (reid_w, reid_h))
    blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
    emb = reid_model([blob])[reid_model.output(0)].flatten()
    return emb / np.linalg.norm(emb)


# Gallery of (numeric_id, embedding). recognize() returns an existing ID for a
# known face or enrolls a new one, keeping each person's ID stable over time.
gallery = []
next_id = 1


def recognize(embedding):
    global next_id
    best_index, best_sim = -1, 0.0
    for index, (_, gallery_emb) in enumerate(gallery):
        sim = float(np.dot(embedding, gallery_emb))
        if sim > best_sim:
            best_sim, best_index = sim, index
    if best_sim >= MATCH_THRESHOLD:
        person_id, gallery_emb = gallery[best_index]
        # Blend the embedding into the gallery entry to stay robust to pose.
        updated = 0.9 * gallery_emb + 0.1 * embedding
        gallery[best_index] = (person_id, updated / np.linalg.norm(updated))
        return person_id
    person_id = next_id
    next_id += 1
    gallery.append((person_id, embedding))
    print(f"Enrolled ID {person_id}")
    return person_id


cap = cv2.VideoCapture(INPUT_VIDEO)
fps = cap.get(cv2.CAP_PROP_FPS) or 12
frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
writer = cv2.VideoWriter(
    "output_openvino.mp4", cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))

while True:
    ok, frame = cap.read()
    if not ok:
        break
    for bbox in detect_faces(frame):
        person_id = recognize(get_embedding(frame, bbox))
        x1, y1, x2, y2 = bbox
        cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv2.putText(frame, f"ID {person_id}", (x1, max(15, y1 - 8)),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
    writer.write(frame)

cap.release()
writer.release()
print(f"Total identities recognized: {len(gallery)}")
print("Saved: output_openvino.mp4")
```

**Device targets:**

- `"CPU"` -- default, works on all Intel platforms.
- `"GPU"` -- Intel integrated or discrete GPU.
- `"NPU"` -- Intel NPU; face-detection-adas-0001 FP16 is NPU-compatible.

#### Expected Output

![OpenVINO expected output](expected_output_openvino.gif)

### DLStreamer Sample

The pipeline below runs the face detector via `gvadetect` and the
re-identification model via `gvaclassify` on the video. Frames are pulled through
an `appsink`, where each face's embedding is matched against a gallery to assign
a stable numeric ID (new people are enrolled, returning people keep their ID).
Every face is annotated with its `ID <n>` and the result is saved to
`output_dlstreamer.mp4`.

> **Notes on running this sample:**
>
> - Export `PYTHONPATH` so the DLStreamer Python modules (`gi`, `gstgva`) are
>   importable:
>
>   ```bash
>   source /opt/intel/openvino_2026/setupvars.sh
>   source /opt/intel/dlstreamer/scripts/setup_dls_env.sh
>   export PYTHONPATH=/opt/intel/dlstreamer/python:\
>   /opt/intel/dlstreamer/gstreamer/lib/python3/dist-packages:${PYTHONPATH:-}
>   ```
>
> - The re-identification embedding is attached as a tensor on each face's
>   region-of-interest metadata. Convert the stream to `BGR` **before**
>   `gvadetect`/`gvaclassify` so a downstream format conversion does not strip
>   those tensors before the `appsink` reads them.

```python
import gi

gi.require_version("Gst", "1.0")
from gi.repository import Gst

Gst.init([])

import numpy as np
import cv2
from gstgva import VideoFrame

INPUT_VIDEO = "test_video.mp4"
OUTPUT_VIDEO = "output_dlstreamer.mp4"
DETECTION_MODEL = "intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml"
REID_MODEL = "intel/face-reidentification-retail-0095/FP16/face-reidentification-retail-0095.xml"
# For CPU: change "GPU" to "CPU". For NPU: change "GPU" to "NPU".
DEVICE = "GPU"
DET_THRESHOLD = 0.6
MATCH_THRESHOLD = 0.5

# Gallery of (numeric_id, embedding). recognize() returns an existing ID for a
# known face or enrolls a new one, keeping each person's ID stable over time.
gallery = []
next_id = 1


def recognize(embedding):
    global next_id
    best_index, best_sim = -1, 0.0
    for index, (_, gallery_emb) in enumerate(gallery):
        sim = float(np.dot(embedding, gallery_emb))
        if sim > best_sim:
            best_sim, best_index = sim, index
    if best_sim >= MATCH_THRESHOLD:
        person_id, gallery_emb = gallery[best_index]
        # Blend the embedding into the gallery entry to stay robust to pose.
        updated = 0.9 * gallery_emb + 0.1 * embedding
        gallery[best_index] = (person_id, updated / np.linalg.norm(updated))
        return person_id
    person_id = next_id
    next_id += 1
    gallery.append((person_id, embedding))
    print(f"Enrolled ID {person_id}", flush=True)
    return person_id


def face_embeddings(video_frame):
    """Yield ((x, y, w, h), normalized_embedding) for each classified face."""
    for region in video_frame.regions():
        rect = region.rect()
        emb = None
        for tensor in region.tensors():
            if tensor.is_detection():
                continue
            data = np.array(tensor.data(), dtype=np.float32)
            if data.size >= 256:
                emb = data[:256]
        if emb is None:
            continue
        emb = emb / (np.linalg.norm(emb) + 1e-9)
        yield (int(rect.x), int(rect.y), int(rect.w), int(rect.h)), emb


# Convert to BGR before inference so gvaclassify's embedding tensors survive to
# the appsink (a later format-changing videoconvert would strip them).
pipeline = Gst.parse_launch(
    f"filesrc location={INPUT_VIDEO} ! decodebin3 ! "
    "videoconvert ! video/x-raw,format=BGR ! "
    f"gvadetect model={DETECTION_MODEL} device={DEVICE} "
    f"threshold={DET_THRESHOLD} ! queue ! "
    f"gvaclassify model={REID_MODEL} device={DEVICE} ! queue ! "
    "appsink name=sink emit-signals=true sync=false max-buffers=4 drop=false"
)
sink = pipeline.get_by_name("sink")

writer = {"w": None}


def on_video(sink):
    sample = sink.emit("pull-sample")
    if sample is None:
        return Gst.FlowReturn.OK
    vf = VideoFrame(sample.get_buffer(), caps=sample.get_caps())
    labeled = []
    for (x, y, w, h), emb in face_embeddings(vf):
        labeled.append((x, y, w, h, recognize(emb)))

    with vf.data() as mat:
        frame = mat.copy()

    if writer["w"] is None:
        frame_h, frame_w = frame.shape[:2]
        structure = sample.get_caps().get_structure(0)
        ok_fr, fps_n, fps_d = structure.get_fraction("framerate")
        fps = fps_n / fps_d if ok_fr and fps_d else 12
        writer["w"] = cv2.VideoWriter(
            OUTPUT_VIDEO, cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))

    for x, y, w, h, person_id in labeled:
        cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.putText(frame, f"ID {person_id}", (x, max(15, y - 8)),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
    writer["w"].write(frame)
    return Gst.FlowReturn.OK


sink.connect("new-sample", on_video)
pipeline.set_state(Gst.State.PLAYING)
pipeline.get_bus().timed_pop_filtered(
    Gst.CLOCK_TIME_NONE, Gst.MessageType.EOS | Gst.MessageType.ERROR)
pipeline.set_state(Gst.State.NULL)
if writer["w"] is not None:
    writer["w"].release()
print(f"Total identities recognized: {len(gallery)}", flush=True)
print(f"Saved: {OUTPUT_VIDEO}", flush=True)
```

**Device targets:**

- `DEVICE = "GPU"` -- default in the sample code.
- `DEVICE = "CPU"` -- change `"GPU"` to `"CPU"`.
- `DEVICE = "NPU"` -- change `"GPU"` to `"NPU"`; use `batch-size=1` and `nireq=4` for best NPU utilization.

#### Expected Output

![DLStreamer expected output](expected_output_dlstreamer.gif)

---

## License

Licensed under the MIT License. See [LICENSE](LICENSE) for details.

## References

- [face-detection-adas-0001](https://docs.openvino.ai/2024/omz_models_model_face_detection_adas_0001.html)
- [face-reidentification-retail-0095](https://docs.openvino.ai/2024/omz_models_model_face_reidentification_retail_0095.html)
- [Open Model Zoo](https://github.com/openvinotoolkit/open_model_zoo)
- [OpenVINO Documentation](https://docs.openvino.ai/)
- [Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/index.html)