vagheshpatel commited on
Commit
75986a7
·
verified ·
1 Parent(s): 9d3bfe0

Update native RC1 samples and supported device documentation

Browse files
Files changed (4) hide show
  1. README.md +26 -2
  2. dlstreamer_sample.py +121 -0
  3. export_and_quantize.sh +19 -19
  4. openvino_sample.py +104 -0
README.md CHANGED
@@ -61,9 +61,13 @@ Typical Metro deployments include:
61
 
62
  ## Prerequisites
63
 
 
 
 
 
64
  - Python 3.11+
65
  - [Install OpenVINO](https://docs.openvino.ai/2026/get-started/install-openvino.html) (latest version)
66
- - [Install Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/get_started/install/install_guide_ubuntu.html) (latest version)
67
 
68
  Create and activate a Python virtual environment before running the scripts:
69
 
@@ -106,7 +110,16 @@ person is seen again, the gallery returns their existing ID. Every face is
106
  annotated with its `ID <n>`, and the result is saved to `output_openvino.mp4`.
107
  Change the `device` string to run on CPU, GPU, or NPU.
108
 
 
 
 
109
  ```python
 
 
 
 
 
 
110
  import cv2
111
  import numpy as np
112
  import openvino as ov
@@ -206,6 +219,7 @@ writer.release()
206
  print(f"Total identities recognized: {len(gallery)}")
207
  print("Saved: output_openvino.mp4")
208
  ```
 
209
 
210
  **Device targets:**
211
 
@@ -243,7 +257,16 @@ Every face is annotated with its `ID <n>` and the result is saved to
243
  > `gvadetect`/`gvaclassify` so a downstream format conversion does not strip
244
  > those tensors before the `appsink` reads them.
245
 
 
 
 
246
  ```python
 
 
 
 
 
 
247
  import gi
248
 
249
  gi.require_version("Gst", "1.0")
@@ -360,6 +383,7 @@ if writer["w"] is not None:
360
  print(f"Total identities recognized: {len(gallery)}", flush=True)
361
  print(f"Saved: {OUTPUT_VIDEO}", flush=True)
362
  ```
 
363
 
364
  **Device targets:**
365
 
@@ -383,4 +407,4 @@ Licensed under the MIT License. See [LICENSE](LICENSE) for details.
383
  - [face-reidentification-retail-0095](https://docs.openvino.ai/2024/omz_models_model_face_reidentification_retail_0095.html)
384
  - [Open Model Zoo](https://github.com/openvinotoolkit/open_model_zoo)
385
  - [OpenVINO Documentation](https://docs.openvino.ai/)
386
- - [Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/index.html)
 
61
 
62
  ## Prerequisites
63
 
64
+ Target stack: Intel DLStreamer **2026.3.0 RC1** with the matching OpenVINO GenAI bundle **2026.5.0.0.dev20261030** (OpenVINO Runtime **2026.5.0.dev20261030**).
65
+ Use the [native RC1 setup](../README.md#native-rc1-runtime) before running these samples; this is a development stack, not a stable release.
66
+ See the [device validation matrix](../DEVICE_SUPPORT_MATRIX.md) for measured CPU, GPU, and NPU results.
67
+
68
  - Python 3.11+
69
  - [Install OpenVINO](https://docs.openvino.ai/2026/get-started/install-openvino.html) (latest version)
70
+ - [Install Intel DLStreamer](https://docs.openedgeplatform.intel.com/dev/edge-ai-libraries/dlstreamer/get_started/install/install_guide_ubuntu.html) (latest version)
71
 
72
  Create and activate a Python virtual environment before running the scripts:
73
 
 
110
  annotated with its `ID <n>`, and the result is saved to `output_openvino.mp4`.
111
  Change the `device` string to run on CPU, GPU, or NPU.
112
 
113
+ The full sample is in [openvino_sample.py](openvino_sample.py); run it with `python3 openvino_sample.py` or copy the code below.
114
+
115
+ <!-- sample: openvino_sample.py -->
116
  ```python
117
+ # SPDX-License-Identifier: MIT
118
+ # Copyright (C) Intel Corporation
119
+ #
120
+ # Facial Recognition: OpenVINO Runtime sample.
121
+ # See README.md for prerequisites and inputs. Run: python3 openvino_sample.py
122
+
123
  import cv2
124
  import numpy as np
125
  import openvino as ov
 
219
  print(f"Total identities recognized: {len(gallery)}")
220
  print("Saved: output_openvino.mp4")
221
  ```
222
+ <!-- /sample -->
223
 
224
  **Device targets:**
225
 
 
257
  > `gvadetect`/`gvaclassify` so a downstream format conversion does not strip
258
  > those tensors before the `appsink` reads them.
259
 
260
+ The full sample is in [dlstreamer_sample.py](dlstreamer_sample.py); run it with `python3 dlstreamer_sample.py` or copy the code below.
261
+
262
+ <!-- sample: dlstreamer_sample.py -->
263
  ```python
264
+ # SPDX-License-Identifier: MIT
265
+ # Copyright (C) Intel Corporation
266
+ #
267
+ # Facial Recognition: DLStreamer pipeline sample.
268
+ # See README.md for prerequisites and inputs. Run: python3 dlstreamer_sample.py
269
+
270
  import gi
271
 
272
  gi.require_version("Gst", "1.0")
 
383
  print(f"Total identities recognized: {len(gallery)}", flush=True)
384
  print(f"Saved: {OUTPUT_VIDEO}", flush=True)
385
  ```
386
+ <!-- /sample -->
387
 
388
  **Device targets:**
389
 
 
407
  - [face-reidentification-retail-0095](https://docs.openvino.ai/2024/omz_models_model_face_reidentification_retail_0095.html)
408
  - [Open Model Zoo](https://github.com/openvinotoolkit/open_model_zoo)
409
  - [OpenVINO Documentation](https://docs.openvino.ai/)
410
+ - [Intel DLStreamer](https://docs.openedgeplatform.intel.com/dev/edge-ai-libraries/dlstreamer/index.html)
dlstreamer_sample.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: MIT
2
+ # Copyright (C) Intel Corporation
3
+ #
4
+ # Facial Recognition: DLStreamer pipeline sample.
5
+ # See README.md for prerequisites and inputs. Run: python3 dlstreamer_sample.py
6
+
7
+ import gi
8
+
9
+ gi.require_version("Gst", "1.0")
10
+ from gi.repository import Gst
11
+
12
+ Gst.init([])
13
+
14
+ import numpy as np
15
+ import cv2
16
+ from gstgva import VideoFrame
17
+
18
+ INPUT_VIDEO = "test_video.mp4"
19
+ OUTPUT_VIDEO = "output_dlstreamer.mp4"
20
+ DETECTION_MODEL = "intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml"
21
+ REID_MODEL = "intel/face-reidentification-retail-0095/FP16/face-reidentification-retail-0095.xml"
22
+ # For CPU: change "GPU" to "CPU". For NPU: change "GPU" to "NPU".
23
+ DEVICE = "GPU"
24
+ DET_THRESHOLD = 0.6
25
+ MATCH_THRESHOLD = 0.5
26
+
27
+ # Gallery of (numeric_id, embedding). recognize() returns an existing ID for a
28
+ # known face or enrolls a new one, keeping each person's ID stable over time.
29
+ gallery = []
30
+ next_id = 1
31
+
32
+
33
+ def recognize(embedding):
34
+ global next_id
35
+ best_index, best_sim = -1, 0.0
36
+ for index, (_, gallery_emb) in enumerate(gallery):
37
+ sim = float(np.dot(embedding, gallery_emb))
38
+ if sim > best_sim:
39
+ best_sim, best_index = sim, index
40
+ if best_sim >= MATCH_THRESHOLD:
41
+ person_id, gallery_emb = gallery[best_index]
42
+ # Blend the embedding into the gallery entry to stay robust to pose.
43
+ updated = 0.9 * gallery_emb + 0.1 * embedding
44
+ gallery[best_index] = (person_id, updated / np.linalg.norm(updated))
45
+ return person_id
46
+ person_id = next_id
47
+ next_id += 1
48
+ gallery.append((person_id, embedding))
49
+ print(f"Enrolled ID {person_id}", flush=True)
50
+ return person_id
51
+
52
+
53
+ def face_embeddings(video_frame):
54
+ """Yield ((x, y, w, h), normalized_embedding) for each classified face."""
55
+ for region in video_frame.regions():
56
+ rect = region.rect()
57
+ emb = None
58
+ for tensor in region.tensors():
59
+ if tensor.is_detection():
60
+ continue
61
+ data = np.array(tensor.data(), dtype=np.float32)
62
+ if data.size >= 256:
63
+ emb = data[:256]
64
+ if emb is None:
65
+ continue
66
+ emb = emb / (np.linalg.norm(emb) + 1e-9)
67
+ yield (int(rect.x), int(rect.y), int(rect.w), int(rect.h)), emb
68
+
69
+
70
+ # Convert to BGR before inference so gvaclassify's embedding tensors survive to
71
+ # the appsink (a later format-changing videoconvert would strip them).
72
+ pipeline = Gst.parse_launch(
73
+ f"filesrc location={INPUT_VIDEO} ! decodebin3 ! "
74
+ "videoconvert ! video/x-raw,format=BGR ! "
75
+ f"gvadetect model={DETECTION_MODEL} device={DEVICE} "
76
+ f"threshold={DET_THRESHOLD} ! queue ! "
77
+ f"gvaclassify model={REID_MODEL} device={DEVICE} ! queue ! "
78
+ "appsink name=sink emit-signals=true sync=false max-buffers=4 drop=false"
79
+ )
80
+ sink = pipeline.get_by_name("sink")
81
+
82
+ writer = {"w": None}
83
+
84
+
85
+ def on_video(sink):
86
+ sample = sink.emit("pull-sample")
87
+ if sample is None:
88
+ return Gst.FlowReturn.OK
89
+ vf = VideoFrame(sample.get_buffer(), caps=sample.get_caps())
90
+ labeled = []
91
+ for (x, y, w, h), emb in face_embeddings(vf):
92
+ labeled.append((x, y, w, h, recognize(emb)))
93
+
94
+ with vf.data() as mat:
95
+ frame = mat.copy()
96
+
97
+ if writer["w"] is None:
98
+ frame_h, frame_w = frame.shape[:2]
99
+ structure = sample.get_caps().get_structure(0)
100
+ ok_fr, fps_n, fps_d = structure.get_fraction("framerate")
101
+ fps = fps_n / fps_d if ok_fr and fps_d else 12
102
+ writer["w"] = cv2.VideoWriter(
103
+ OUTPUT_VIDEO, cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))
104
+
105
+ for x, y, w, h, person_id in labeled:
106
+ cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
107
+ cv2.putText(frame, f"ID {person_id}", (x, max(15, y - 8)),
108
+ cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
109
+ writer["w"].write(frame)
110
+ return Gst.FlowReturn.OK
111
+
112
+
113
+ sink.connect("new-sample", on_video)
114
+ pipeline.set_state(Gst.State.PLAYING)
115
+ pipeline.get_bus().timed_pop_filtered(
116
+ Gst.CLOCK_TIME_NONE, Gst.MessageType.EOS | Gst.MessageType.ERROR)
117
+ pipeline.set_state(Gst.State.NULL)
118
+ if writer["w"] is not None:
119
+ writer["w"].release()
120
+ print(f"Total identities recognized: {len(gallery)}", flush=True)
121
+ print(f"Saved: {OUTPUT_VIDEO}", flush=True)
export_and_quantize.sh CHANGED
@@ -11,25 +11,6 @@ set -euo pipefail
11
  # Official Open Model Zoo public model storage (versioned, immutable).
12
  OMZ_BASE="https://storage.openvinotoolkit.org/repositories/open_model_zoo/2023.0/models_bin/1"
13
 
14
- echo "--- Installing dependencies ---"
15
- pip install -qU openvino
16
-
17
- # Download both the IR topology (.xml) and weights (.bin) for an OMZ model
18
- # from the official storage into intel/<model>/<precision>/.
19
- download_omz_model() {
20
- local model="$1"
21
- local precision="$2"
22
- local dest="intel/${model}/${precision}"
23
- mkdir -p "${dest}"
24
- local ext
25
- for ext in xml bin; do
26
- if [[ ! -f "${dest}/${model}.${ext}" ]]; then
27
- wget -q -O "${dest}/${model}.${ext}" \
28
- "${OMZ_BASE}/${model}/${precision}/${model}.${ext}"
29
- fi
30
- done
31
- }
32
-
33
  # Ask for approval before downloading models and sample files
34
  echo ""
35
  echo "This script will download:"
@@ -53,6 +34,25 @@ else
53
  echo "WARNING: HuggingFace tracking ping failed (offline?); continuing." >&2
54
  fi
55
  echo ""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
56
 
57
  echo "--- Downloading face-detection-adas-0001 (FP16) ---"
58
  download_omz_model face-detection-adas-0001 FP16
 
11
  # Official Open Model Zoo public model storage (versioned, immutable).
12
  OMZ_BASE="https://storage.openvinotoolkit.org/repositories/open_model_zoo/2023.0/models_bin/1"
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  # Ask for approval before downloading models and sample files
15
  echo ""
16
  echo "This script will download:"
 
34
  echo "WARNING: HuggingFace tracking ping failed (offline?); continuing." >&2
35
  fi
36
  echo ""
37
+ echo "--- Installing dependencies ---"
38
+ pip install -qU openvino
39
+
40
+ # Download both the IR topology (.xml) and weights (.bin) for an OMZ model
41
+ # from the official storage into intel/<model>/<precision>/.
42
+ download_omz_model() {
43
+ local model="$1"
44
+ local precision="$2"
45
+ local dest="intel/${model}/${precision}"
46
+ mkdir -p "${dest}"
47
+ local ext
48
+ for ext in xml bin; do
49
+ if [[ ! -f "${dest}/${model}.${ext}" ]]; then
50
+ wget -q -O "${dest}/${model}.${ext}" \
51
+ "${OMZ_BASE}/${model}/${precision}/${model}.${ext}"
52
+ fi
53
+ done
54
+ }
55
+
56
 
57
  echo "--- Downloading face-detection-adas-0001 (FP16) ---"
58
  download_omz_model face-detection-adas-0001 FP16
openvino_sample.py ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: MIT
2
+ # Copyright (C) Intel Corporation
3
+ #
4
+ # Facial Recognition: OpenVINO Runtime sample.
5
+ # See README.md for prerequisites and inputs. Run: python3 openvino_sample.py
6
+
7
+ import cv2
8
+ import numpy as np
9
+ import openvino as ov
10
+
11
+ DETECTION_MODEL = "intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml"
12
+ REID_MODEL = "intel/face-reidentification-retail-0095/FP16/face-reidentification-retail-0095.xml"
13
+ INPUT_VIDEO = "test_video.mp4"
14
+ CONF_THRESHOLD = 0.6
15
+ MATCH_THRESHOLD = 0.5
16
+
17
+ core = ov.Core()
18
+
19
+ # Change device to "GPU" or "NPU" to run on integrated GPU or NPU.
20
+ det_model = core.compile_model(core.read_model(DETECTION_MODEL), "CPU")
21
+ reid_model = core.compile_model(core.read_model(REID_MODEL), "CPU")
22
+
23
+ det_input = det_model.input(0)
24
+ det_h, det_w = det_input.shape[2], det_input.shape[3]
25
+ reid_input = reid_model.input(0)
26
+ reid_h, reid_w = reid_input.shape[2], reid_input.shape[3]
27
+
28
+
29
+ def detect_faces(img):
30
+ h0, w0 = img.shape[:2]
31
+ blob = cv2.resize(img, (det_w, det_h))
32
+ blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
33
+ detections = det_model([blob])[det_model.output(0)][0][0]
34
+ boxes = []
35
+ for det in detections:
36
+ if float(det[2]) < CONF_THRESHOLD:
37
+ continue
38
+ x1 = max(0, int(det[3] * w0))
39
+ y1 = max(0, int(det[4] * h0))
40
+ x2 = min(w0, int(det[5] * w0))
41
+ y2 = min(h0, int(det[6] * h0))
42
+ if x2 > x1 and y2 > y1:
43
+ boxes.append((x1, y1, x2, y2))
44
+ return boxes
45
+
46
+
47
+ def get_embedding(img, bbox):
48
+ x1, y1, x2, y2 = bbox
49
+ crop = img[y1:y2, x1:x2]
50
+ blob = cv2.resize(crop, (reid_w, reid_h))
51
+ blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
52
+ emb = reid_model([blob])[reid_model.output(0)].flatten()
53
+ return emb / np.linalg.norm(emb)
54
+
55
+
56
+ # Gallery of (numeric_id, embedding). recognize() returns an existing ID for a
57
+ # known face or enrolls a new one, keeping each person's ID stable over time.
58
+ gallery = []
59
+ next_id = 1
60
+
61
+
62
+ def recognize(embedding):
63
+ global next_id
64
+ best_index, best_sim = -1, 0.0
65
+ for index, (_, gallery_emb) in enumerate(gallery):
66
+ sim = float(np.dot(embedding, gallery_emb))
67
+ if sim > best_sim:
68
+ best_sim, best_index = sim, index
69
+ if best_sim >= MATCH_THRESHOLD:
70
+ person_id, gallery_emb = gallery[best_index]
71
+ # Blend the embedding into the gallery entry to stay robust to pose.
72
+ updated = 0.9 * gallery_emb + 0.1 * embedding
73
+ gallery[best_index] = (person_id, updated / np.linalg.norm(updated))
74
+ return person_id
75
+ person_id = next_id
76
+ next_id += 1
77
+ gallery.append((person_id, embedding))
78
+ print(f"Enrolled ID {person_id}")
79
+ return person_id
80
+
81
+
82
+ cap = cv2.VideoCapture(INPUT_VIDEO)
83
+ fps = cap.get(cv2.CAP_PROP_FPS) or 12
84
+ frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
85
+ frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
86
+ writer = cv2.VideoWriter(
87
+ "output_openvino.mp4", cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))
88
+
89
+ while True:
90
+ ok, frame = cap.read()
91
+ if not ok:
92
+ break
93
+ for bbox in detect_faces(frame):
94
+ person_id = recognize(get_embedding(frame, bbox))
95
+ x1, y1, x2, y2 = bbox
96
+ cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
97
+ cv2.putText(frame, f"ID {person_id}", (x1, max(15, y1 - 8)),
98
+ cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
99
+ writer.write(frame)
100
+
101
+ cap.release()
102
+ writer.release()
103
+ print(f"Total identities recognized: {len(gallery)}")
104
+ print("Saved: output_openvino.mp4")