vagheshpatel commited on
Commit
274565f
·
verified ·
1 Parent(s): ade60cb

Update native RC1 samples and supported device documentation

Browse files
Files changed (4) hide show
  1. README.md +26 -2
  2. dlstreamer_sample.py +117 -0
  3. export_and_quantize.sh +19 -19
  4. openvino_sample.py +103 -0
README.md CHANGED
@@ -80,9 +80,13 @@ Typical Metro deployments include:
80
 
81
  ## Prerequisites
82
 
 
 
 
 
83
  - Python 3.11+
84
  - [Install OpenVINO](https://docs.openvino.ai/2026/get-started/install-openvino.html) (latest version)
85
- - [Install Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/get_started/install/install_guide_ubuntu.html) (latest version)
86
 
87
  Create and activate a Python virtual environment before running the scripts:
88
 
@@ -122,7 +126,16 @@ query. It saves the frame containing the most matching people, drawing a green
122
  box on every matched person.
123
  Change the `device` string to run on CPU, GPU, or NPU.
124
 
 
 
 
125
  ```python
 
 
 
 
 
 
126
  import cv2
127
  import numpy as np
128
  import openvino as ov
@@ -221,6 +234,7 @@ print(f"Found {best['count']} person(s) matching [{query_text}]")
221
  cv2.imwrite("output_openvino.jpg", frame)
222
  print("Saved: output_openvino.jpg")
223
  ```
 
224
 
225
  **Device targets:**
226
 
@@ -259,7 +273,16 @@ Only people whose appearance matches the query are boxed, so the annotated
259
  > `gvadetect`/`gvaclassify` so a downstream format conversion does not strip
260
  > those tensors before the `appsink` reads them.
261
 
 
 
 
262
  ```python
 
 
 
 
 
 
263
  import gi
264
 
265
  gi.require_version("Gst", "1.0")
@@ -372,6 +395,7 @@ if writer["w"] is not None:
372
  print(f"Frames with a person matching [{query_text}]: {match_frames}", flush=True)
373
  print(f"Saved: {OUTPUT_VIDEO}", flush=True)
374
  ```
 
375
 
376
  **Device targets:**
377
 
@@ -395,4 +419,4 @@ Licensed under the MIT License. See [LICENSE](LICENSE) for details.
395
  - [person-attributes-recognition-crossroad-0230](https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/intel/person-attributes-recognition-crossroad-0230/README.md)
396
  - [Open Model Zoo](https://github.com/openvinotoolkit/open_model_zoo)
397
  - [OpenVINO Documentation](https://docs.openvino.ai/)
398
- - [Intel DLStreamer](https://docs.openedgeplatform.intel.com/2026.0/edge-ai-libraries/dlstreamer/index.html)
 
80
 
81
  ## Prerequisites
82
 
83
+ 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**).
84
+ Use the [native RC1 setup](../README.md#native-rc1-runtime) before running these samples; this is a development stack, not a stable release.
85
+ See the [device validation matrix](../DEVICE_SUPPORT_MATRIX.md) for measured CPU, GPU, and NPU results.
86
+
87
  - Python 3.11+
88
  - [Install OpenVINO](https://docs.openvino.ai/2026/get-started/install-openvino.html) (latest version)
89
+ - [Install Intel DLStreamer](https://docs.openedgeplatform.intel.com/dev/edge-ai-libraries/dlstreamer/get_started/install/install_guide_ubuntu.html) (latest version)
90
 
91
  Create and activate a Python virtual environment before running the scripts:
92
 
 
126
  box on every matched person.
127
  Change the `device` string to run on CPU, GPU, or NPU.
128
 
129
+ The full sample is in [openvino_sample.py](openvino_sample.py); run it with `python3 openvino_sample.py` or copy the code below.
130
+
131
+ <!-- sample: openvino_sample.py -->
132
  ```python
133
+ # SPDX-License-Identifier: MIT
134
+ # Copyright (C) Intel Corporation
135
+ #
136
+ # Appearance-Based Search: OpenVINO Runtime sample.
137
+ # See README.md for prerequisites and inputs. Run: python3 openvino_sample.py
138
+
139
  import cv2
140
  import numpy as np
141
  import openvino as ov
 
234
  cv2.imwrite("output_openvino.jpg", frame)
235
  print("Saved: output_openvino.jpg")
236
  ```
237
+ <!-- /sample -->
238
 
239
  **Device targets:**
240
 
 
273
  > `gvadetect`/`gvaclassify` so a downstream format conversion does not strip
274
  > those tensors before the `appsink` reads them.
275
 
276
+ The full sample is in [dlstreamer_sample.py](dlstreamer_sample.py); run it with `python3 dlstreamer_sample.py` or copy the code below.
277
+
278
+ <!-- sample: dlstreamer_sample.py -->
279
  ```python
280
+ # SPDX-License-Identifier: MIT
281
+ # Copyright (C) Intel Corporation
282
+ #
283
+ # Appearance-Based Search: DLStreamer pipeline sample.
284
+ # See README.md for prerequisites and inputs. Run: python3 dlstreamer_sample.py
285
+
286
  import gi
287
 
288
  gi.require_version("Gst", "1.0")
 
395
  print(f"Frames with a person matching [{query_text}]: {match_frames}", flush=True)
396
  print(f"Saved: {OUTPUT_VIDEO}", flush=True)
397
  ```
398
+ <!-- /sample -->
399
 
400
  **Device targets:**
401
 
 
419
  - [person-attributes-recognition-crossroad-0230](https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/intel/person-attributes-recognition-crossroad-0230/README.md)
420
  - [Open Model Zoo](https://github.com/openvinotoolkit/open_model_zoo)
421
  - [OpenVINO Documentation](https://docs.openvino.ai/)
422
+ - [Intel DLStreamer](https://docs.openedgeplatform.intel.com/dev/edge-ai-libraries/dlstreamer/index.html)
dlstreamer_sample.py ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: MIT
2
+ # Copyright (C) Intel Corporation
3
+ #
4
+ # Appearance-Based Search: 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/person-detection-retail-0013/FP16/person-detection-retail-0013.xml"
21
+ ATTRIBUTES_MODEL = "intel/person-attributes-recognition-crossroad-0230/FP16/person-attributes-recognition-crossroad-0230.xml"
22
+ # For CPU: change "GPU" to "CPU". For NPU: change "GPU" to "NPU".
23
+ DEVICE = "GPU"
24
+ DET_THRESHOLD = 0.6
25
+ ATTR_THRESHOLD = 0.5
26
+
27
+ # person-attributes-recognition-crossroad-0230 emits eight binary appearance
28
+ # attributes (output layer "453"), in this order:
29
+ ATTRIBUTE_NAMES = [
30
+ "is_male", "has_bag", "has_backpack", "has_hat",
31
+ "has_longsleeves", "has_longpants", "has_longhair", "has_coat_jacket",
32
+ ]
33
+
34
+ # The appearance being searched for. Set an attribute to True to require it or
35
+ # False to require its absence; omit attributes you do not care about.
36
+ QUERY = {"has_longhair": True}
37
+
38
+
39
+ def person_attributes(video_frame):
40
+ """Yield ((x, y, w, h), {attribute: score}) for each classified person."""
41
+ for region in video_frame.regions():
42
+ rect = region.rect()
43
+ scores = None
44
+ for tensor in region.tensors():
45
+ if tensor.is_detection():
46
+ continue
47
+ data = np.array(tensor.data(), dtype=np.float32)
48
+ # The attributes vector is the length-8 output ("453"); the model
49
+ # also emits two length-2 colour points that are ignored here.
50
+ if data.size == len(ATTRIBUTE_NAMES):
51
+ scores = {name: float(data[i])
52
+ for i, name in enumerate(ATTRIBUTE_NAMES)}
53
+ if scores is None:
54
+ continue
55
+ yield (int(rect.x), int(rect.y), int(rect.w), int(rect.h)), scores
56
+
57
+
58
+ def matches_query(scores):
59
+ return all((scores[name] >= ATTR_THRESHOLD) == wanted
60
+ for name, wanted in QUERY.items())
61
+
62
+
63
+ query_text = ", ".join(k if v else f"no {k}" for k, v in QUERY.items())
64
+ writer = {"w": None}
65
+ match_frames = 0
66
+
67
+
68
+ def on_video(sink):
69
+ global match_frames
70
+ sample = sink.emit("pull-sample")
71
+ if sample is None:
72
+ return Gst.FlowReturn.OK
73
+ vf = VideoFrame(sample.get_buffer(), caps=sample.get_caps())
74
+ matches = [(x, y, w, h) for (x, y, w, h), scores in person_attributes(vf)
75
+ if matches_query(scores)]
76
+
77
+ with vf.data() as mat:
78
+ frame = mat.copy()
79
+
80
+ if writer["w"] is None:
81
+ frame_h, frame_w = frame.shape[:2]
82
+ structure = sample.get_caps().get_structure(0)
83
+ ok_fr, fps_n, fps_d = structure.get_fraction("framerate")
84
+ fps = fps_n / fps_d if ok_fr and fps_d else 12
85
+ writer["w"] = cv2.VideoWriter(
86
+ OUTPUT_VIDEO, cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))
87
+
88
+ for x, y, w, h in matches:
89
+ cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
90
+ cv2.putText(frame, f"MATCH: {query_text}", (x, max(15, y - 8)),
91
+ cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
92
+ if matches:
93
+ match_frames += 1
94
+ writer["w"].write(frame)
95
+ return Gst.FlowReturn.OK
96
+
97
+
98
+ # Convert to BGR before inference so gvaclassify's attribute tensors survive to
99
+ # the appsink (a later format-changing videoconvert would strip them).
100
+ pipeline = Gst.parse_launch(
101
+ f"filesrc location={INPUT_VIDEO} ! decodebin3 ! videoconvert ! "
102
+ f"video/x-raw,format=BGR ! "
103
+ f"gvadetect model={DETECTION_MODEL} device={DEVICE} "
104
+ f"threshold={DET_THRESHOLD} ! queue ! "
105
+ f"gvaclassify model={ATTRIBUTES_MODEL} device={DEVICE} ! queue ! "
106
+ "appsink name=sink emit-signals=true sync=false max-buffers=4 drop=false"
107
+ )
108
+ sink = pipeline.get_by_name("sink")
109
+ sink.connect("new-sample", on_video)
110
+ pipeline.set_state(Gst.State.PLAYING)
111
+ pipeline.get_bus().timed_pop_filtered(
112
+ Gst.CLOCK_TIME_NONE, Gst.MessageType.EOS | Gst.MessageType.ERROR)
113
+ pipeline.set_state(Gst.State.NULL)
114
+ if writer["w"] is not None:
115
+ writer["w"].release()
116
+ print(f"Frames with a person matching [{query_text}]: {match_frames}", flush=True)
117
+ print(f"Saved: {OUTPUT_VIDEO}", flush=True)
export_and_quantize.sh CHANGED
@@ -13,25 +13,6 @@ set -euo pipefail
13
  # Official Open Model Zoo public model storage (versioned, immutable).
14
  OMZ_BASE="https://storage.openvinotoolkit.org/repositories/open_model_zoo/2023.0/models_bin/1"
15
 
16
- echo "--- Installing dependencies ---"
17
- pip install -qU openvino opencv-python numpy
18
-
19
- # Download both the IR topology (.xml) and weights (.bin) for an OMZ model
20
- # from the official storage into intel/<model>/<precision>/.
21
- download_omz_model() {
22
- local model="$1"
23
- local precision="$2"
24
- local dest="intel/${model}/${precision}"
25
- mkdir -p "${dest}"
26
- local ext
27
- for ext in xml bin; do
28
- if [[ ! -f "${dest}/${model}.${ext}" ]]; then
29
- wget -q -O "${dest}/${model}.${ext}" \
30
- "${OMZ_BASE}/${model}/${precision}/${model}.${ext}"
31
- fi
32
- done
33
- }
34
-
35
  # Ask for approval before downloading models and sample files
36
  echo ""
37
  echo "This script will download:"
@@ -55,6 +36,25 @@ else
55
  echo "WARNING: HuggingFace tracking ping failed (offline?); continuing." >&2
56
  fi
57
  echo ""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
 
59
  echo "--- Downloading person-detection-retail-0013 (FP16) ---"
60
  download_omz_model person-detection-retail-0013 FP16
 
13
  # Official Open Model Zoo public model storage (versioned, immutable).
14
  OMZ_BASE="https://storage.openvinotoolkit.org/repositories/open_model_zoo/2023.0/models_bin/1"
15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  # Ask for approval before downloading models and sample files
17
  echo ""
18
  echo "This script will download:"
 
36
  echo "WARNING: HuggingFace tracking ping failed (offline?); continuing." >&2
37
  fi
38
  echo ""
39
+ echo "--- Installing dependencies ---"
40
+ pip install -q openvino opencv-python
41
+
42
+ # Download both the IR topology (.xml) and weights (.bin) for an OMZ model
43
+ # from the official storage into intel/<model>/<precision>/.
44
+ download_omz_model() {
45
+ local model="$1"
46
+ local precision="$2"
47
+ local dest="intel/${model}/${precision}"
48
+ mkdir -p "${dest}"
49
+ local ext
50
+ for ext in xml bin; do
51
+ if [[ ! -f "${dest}/${model}.${ext}" ]]; then
52
+ wget -q -O "${dest}/${model}.${ext}" \
53
+ "${OMZ_BASE}/${model}/${precision}/${model}.${ext}"
54
+ fi
55
+ done
56
+ }
57
+
58
 
59
  echo "--- Downloading person-detection-retail-0013 (FP16) ---"
60
  download_omz_model person-detection-retail-0013 FP16
openvino_sample.py ADDED
@@ -0,0 +1,103 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: MIT
2
+ # Copyright (C) Intel Corporation
3
+ #
4
+ # Appearance-Based Search: 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/person-detection-retail-0013/FP16/person-detection-retail-0013.xml"
12
+ ATTRIBUTES_MODEL = "intel/person-attributes-recognition-crossroad-0230/FP16/person-attributes-recognition-crossroad-0230.xml"
13
+ SCENE_VIDEO = "test_video.mp4"
14
+
15
+ # person-attributes-recognition-crossroad-0230 emits eight binary appearance
16
+ # attributes (output layer "453"), in this order:
17
+ ATTRIBUTE_NAMES = [
18
+ "is_male", "has_bag", "has_backpack", "has_hat",
19
+ "has_longsleeves", "has_longpants", "has_longhair", "has_coat_jacket",
20
+ ]
21
+
22
+ # The appearance being searched for. Set an attribute to True to require it or
23
+ # False to require its absence; omit attributes you do not care about.
24
+ QUERY = {"has_longhair": True}
25
+
26
+ CONF_THRESHOLD = 0.6 # person-detection confidence
27
+ ATTR_THRESHOLD = 0.5 # attribute presence threshold
28
+
29
+ core = ov.Core()
30
+
31
+ # Change device to "GPU" or "NPU" to run on integrated GPU or NPU.
32
+ det_compiled = core.compile_model(core.read_model(DETECTION_MODEL), "CPU")
33
+ attr_compiled = core.compile_model(core.read_model(ATTRIBUTES_MODEL), "CPU")
34
+
35
+ det_input = det_compiled.input(0)
36
+ det_h, det_w = det_input.shape[2], det_input.shape[3]
37
+ attr_input = attr_compiled.input(0)
38
+ attr_h, attr_w = attr_input.shape[2], attr_input.shape[3]
39
+ attr_output = attr_compiled.output("453")
40
+
41
+
42
+ def detect_persons(img):
43
+ h0, w0 = img.shape[:2]
44
+ blob = cv2.resize(img, (det_w, det_h))
45
+ blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
46
+ dets = det_compiled([blob])[det_compiled.output(0)][0][0]
47
+ persons = []
48
+ for d in dets:
49
+ if float(d[2]) < CONF_THRESHOLD:
50
+ continue
51
+ x1 = max(0, int(d[3] * w0))
52
+ y1 = max(0, int(d[4] * h0))
53
+ x2 = min(w0, int(d[5] * w0))
54
+ y2 = min(h0, int(d[6] * h0))
55
+ if x2 > x1 and y2 > y1:
56
+ persons.append((x1, y1, x2, y2))
57
+ return persons
58
+
59
+
60
+ def get_attributes(img, bbox):
61
+ x1, y1, x2, y2 = bbox
62
+ crop = img[y1:y2, x1:x2]
63
+ blob = cv2.resize(crop, (attr_w, attr_h))
64
+ blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
65
+ values = attr_compiled([blob])[attr_output].flatten()
66
+ return {name: float(values[i]) for i, name in enumerate(ATTRIBUTE_NAMES)}
67
+
68
+
69
+ def matches_query(attributes):
70
+ return all((attributes[name] >= ATTR_THRESHOLD) == wanted
71
+ for name, wanted in QUERY.items())
72
+
73
+
74
+ # Scan the scene and keep the frame containing the most people whose appearance
75
+ # matches the query, so the saved image best illustrates the search result.
76
+ cap = cv2.VideoCapture(SCENE_VIDEO)
77
+ total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 900
78
+ best = {"count": 0, "frame": None, "boxes": []}
79
+ for frame_idx in range(0, total, 15):
80
+ cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
81
+ ok, frame = cap.read()
82
+ if not ok:
83
+ break
84
+ boxes = [bbox for bbox in detect_persons(frame)
85
+ if matches_query(get_attributes(frame, bbox))]
86
+ if len(boxes) > best["count"]:
87
+ best = {"count": len(boxes), "frame": frame.copy(), "boxes": boxes}
88
+ cap.release()
89
+
90
+ if best["frame"] is None or best["count"] == 0:
91
+ raise SystemExit("No person matching the appearance query was found")
92
+
93
+ # Draw a green box on every person that matches the searched appearance.
94
+ query_text = ", ".join(k if v else f"no {k}" for k, v in QUERY.items())
95
+ frame = best["frame"]
96
+ for x1, y1, x2, y2 in best["boxes"]:
97
+ cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
98
+ cv2.putText(frame, f"MATCH: {query_text}", (x1, max(15, y1 - 8)),
99
+ cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
100
+ print(f"Found {best['count']} person(s) matching [{query_text}]")
101
+
102
+ cv2.imwrite("output_openvino.jpg", frame)
103
+ print("Saved: output_openvino.jpg")