This demo showcases inference of Object Detection networks using Sync and Async API.
Async API usage can improve overall frame-rate of the application, because rather than wait for inference to complete,
the app can continue doing things on the host, while accelerator is busy.
Specifically, this demo keeps the number of Infer Requests that you have set using -nireq flag.
While some of the Infer Requests are processed by IE, the other ones can be filled with new frame data
and asynchronously started or the next output can be taken from the Infer Request and displayed.
This technique can be generalized to any available parallel slack, for example, doing inference and simultaneously encoding the resulting (previous) frames or running further inference, like some emotion detection on top of the face detection results. There are important performance caveats though, for example the tasks that run in parallel should try to avoid oversubscribing the shared compute resources. As another example, if the inference is performed on the HDDL, and the CPU is essentially idle, then it makes sense to do things on the CPU in parallel. But if the inference is performed say on the GPU, then there is little gain from doing the (resulting video) encoding on the same GPU in parallel, because the device is already busy.
This and other performance implications and tips for the Async API are covered in the Optimization Guide.
Other demo objectives are:
- Video as input support via OpenCV*
- Visualization of the resulting bounding boxes and text labels (from the
.labelsfile) or class number (if no file is provided)
On startup, the application reads command-line parameters and loads a network to the Inference Engine. Upon getting a frame from the OpenCV VideoCapture, it performs inference and displays the results.
Async API operates with a notion of the "Infer Request" that encapsulates the inputs/outputs and separates scheduling and waiting for result.
NOTE: By default, Open Model Zoo demos expect input with BGR channels order. If you trained your model to work with RGB order, you need to manually rearrange the default channels order in the demo application or reconvert your model using the Model Optimizer tool with the
--reverse_input_channelsargument specified. For more information about the argument, refer to When to Reverse Input Channels section of Converting a Model Using General Conversion Parameters.
For demo input image or video files, refer to the section Media Files Available for Demos in the Open Model Zoo Demos Overview.
The list of models supported by the demo is in <omz_dir>/demos/object_detection_demo/python/models.lst file.
This file can be used as a parameter for Model Downloader and Converter to download and, if necessary, convert models to OpenVINO Inference Engine format (*.xml + *.bin).
An example of using the Model Downloader:
omz_downloader --list models.lstAn example of using the Model Converter:
omz_converter --list models.lst- architecture_type = centernet
- ctdet_coco_dlav0_512
- architecture_type = ctpn
- ctpn
- architecture_type = detr
- detr-resnet50
- architecture_type = faceboxes
- faceboxes-pytorch
- architecture_type = retinaface-pytorch
- retinaface-resnet50-pytorch
- architecture_type = ssd
- efficientdet-d0-tf
- efficientdet-d1-tf
- face-detection-0200
- face-detection-0202
- face-detection-0204
- face-detection-0205
- face-detection-0206
- face-detection-adas-0001
- face-detection-retail-0004
- face-detection-retail-0005
- face-detection-retail-0044
- faster-rcnn-resnet101-coco-sparse-60-0001
- pedestrian-and-vehicle-detector-adas-0001
- pedestrian-detection-adas-0002
- pelee-coco
- person-detection-0106
- person-detection-0200
- person-detection-0201
- person-detection-0202
- person-detection-0203
- person-detection-0301
- person-detection-0302
- person-detection-0303
- person-detection-retail-0013
- person-vehicle-bike-detection-2000
- person-vehicle-bike-detection-2001
- person-vehicle-bike-detection-2002
- person-vehicle-bike-detection-2003
- person-vehicle-bike-detection-2004
- product-detection-0001
- retinanet-tf
- rfcn-resnet101-coco-tf
- ssd300
- ssd512
- ssd_mobilenet_v1_coco
- ssd_mobilenet_v1_fpn_coco
- ssd-resnet34-1200-onnx
- ssdlite_mobilenet_v2
- vehicle-detection-0200
- vehicle-detection-0201
- vehicle-detection-0202
- vehicle-detection-adas-0002
- vehicle-license-plate-detection-barrier-0106
- vehicle-license-plate-detection-barrier-0123
- architecture_type = ultra_lightweight_face_detection
- ultra-lightweight-face-detection-rfb-320
- ultra-lightweight-face-detection-slim-320
- architecture_type = yolo
- mobilefacedet-v1-mxnet
- mobilenet-yolo-v4-syg
- person-vehicle-bike-detection-crossroad-yolov3-1020
- yolo-v1-tiny-tf
- yolo-v2-ava-0001
- yolo-v2-ava-sparse-35-0001
- yolo-v2-ava-sparse-70-0001
- yolo-v2-tf
- yolo-v2-tiny-ava-0001
- yolo-v2-tiny-ava-sparse-30-0001
- yolo-v2-tiny-ava-sparse-60-0001
- yolo-v2-tiny-tf
- yolo-v2-tiny-vehicle-detection-0001
- yolo-v3-tf
- yolo-v3-tiny-tf
- architecture_type = yolov3-onnx
- yolo-v3-onnx
- yolo-v3-tiny-onnx
- architecture_type = yolov4
- yolo-v4-tf
- yolo-v4-tiny-tf
- architecture_type = yolof
- yolof
- architecture_type = yolox
- yolox-tiny
NOTE: Refer to the tables Intel's Pre-Trained Models Device Support and Public Pre-Trained Models Device Support for the details on models inference support at different devices.
Running the application with the -h option yields the following usage message:
usage: object_detection_demo.py [-h] -m MODEL -at
{ssd,yolo,yolov3-onnx,yolov4,yolof,yolox,faceboxes,centernet,ctpn,retinaface,ultra_lightweight_face_detection,retinaface-pytorch,detr}
-i INPUT [--adapter {openvino,ovms}]
[-d DEVICE] [--labels LABELS] [-t PROB_THRESHOLD]
[--resize_type {standard,fit_to_window,fit_to_window_letterbox}]
[--input_size INPUT_SIZE INPUT_SIZE] [--anchors ANCHORS [ANCHORS ...]]
[--masks MASKS [MASKS ...]] [-nireq NUM_INFER_REQUESTS] [-nstreams NUM_STREAMS]
[-nthreads NUM_THREADS] [--loop] [-o OUTPUT] [-limit OUTPUT_LIMIT] [--no_show]
[--output_resolution OUTPUT_RESOLUTION] [-u UTILIZATION_MONITORS]
[--reverse_input_channels] [--mean_values MEAN_VALUES MEAN_VALUES MEAN_VALUES]
[--scale_values SCALE_VALUES SCALE_VALUES SCALE_VALUES] [-r]
Options:
-h, --help Show this help message and exit.
-m MODEL, --model MODEL
Required. Path to an .xml file with a trained model or
address of model inference service if using OVMS adapter.
-at, --architecture_type Required. Specify model' architecture type. Valid values are {ssd,yolo,yolov3-onnx,yolov4,yolof,yolox,faceboxes,centernet,ctpn,retinaface,ultra_lightweight_face_detection,retinaface-pytorch,detr}.
-i INPUT, --input INPUT
Required. An input to process. The input must be a
single image, a folder of images, video file or camera id.
--adapter {openvino,ovms}
Optional. Specify the model adapter. Default is
openvino.
-d DEVICE, --device DEVICE
Optional. Specify the target device to infer on; CPU,
GPU, HDDL or MYRIAD is acceptable. The demo
will look for a suitable plugin for device specified.
Default value is CPU.
Common model options:
--labels LABELS Optional. Labels mapping file.
-t PROB_THRESHOLD, --prob_threshold PROB_THRESHOLD
Optional. Probability threshold for detections
filtering.
--resize_type {standard,fit_to_window,fit_to_window_letterbox}
Optional. A resize type for model preprocess. By defauld used model predefined type.
--input_size INPUT_SIZE INPUT_SIZE
Optional. The first image size used for CTPN model
reshaping. Default: 600 600. Note that submitted
images should have the same resolution, otherwise
predictions might be incorrect.
--anchors ANCHORS [ANCHORS ...]
Optional. A space separated list of anchors. By default used default anchors for model. Only
for YOLOV4 architecture type.
--masks MASKS [MASKS ...]
Optional. A space separated list of mask for anchors. By default used default masks for model.
Only for YOLOV4 architecture type.
Inference options:
-nireq NUM_INFER_REQUESTS, --num_infer_requests NUM_INFER_REQUESTS
Optional. Number of infer requests
-nstreams NUM_STREAMS, --num_streams NUM_STREAMS
Optional. Number of streams to use for inference on
the CPU or/and GPU in throughput mode (for HETERO and
MULTI device cases use format
<device1>:<nstreams1>,<device2>:<nstreams2> or just
<nstreams>).
-nthreads NUM_THREADS, --num_threads NUM_THREADS
Optional. Number of threads to use for inference on
CPU (including HETERO cases).
Input/output options:
--loop Optional. Enable reading the input in a loop.
-o OUTPUT, --output OUTPUT
Optional. Name of the output file(s) to save.
-limit OUTPUT_LIMIT, --output_limit OUTPUT_LIMIT
Optional. Number of frames to store in output.
If 0 is set, all frames are stored.
--no_show Optional. Don't show output.
--output_resolution OUTPUT_RESOLUTION
Optional. Specify the maximum output window resolution
in (width x height) format. Example: 1280x720.
Input frame size used by default.
-u UTILIZATION_MONITORS, --utilization_monitors UTILIZATION_MONITORS
Optional. List of monitors to show initially.
Input transform options:
--reverse_input_channels REVERSE_CHANNELS
Optional. Switch the input channels order from
BGR to RGB.
--mean_values MEAN_VALUES
Optional. Normalize input by subtracting the mean
values per channel. Example: 255.0 255.0 255.0
--scale_values SCALE_VALUES
Optional. Divide input by scale values per channel.
Division is applied after mean values subtraction.
Example: 255.0 255.0 255.0
Debug options:
-r, --raw_output_message
Optional. Output inference results raw values showing.
Running the application with the empty list of options yields the usage message given above and an error message.
You can use the following command to do inference on GPU with a pre-trained object detection model:
python3 object_detection_demo.py \
-d GPU \
-i <path_to_video>/inputVideo.mp4 \
-m <path_to_model>/ssd300.xml \
-at ssd \
--labels <omz_dir>/data/dataset_classes/voc_20cl_bkgr.txtThe number of Infer Requests is specified by -nireq flag. An increase of this number usually leads to an increase
of performance (throughput), since in this case several Infer Requests can be processed simultaneously if the device
supports parallelization. However, a large number of Infer Requests increases the latency because each frame still
has to wait before being sent for inference.
For higher FPS, it is recommended that you set -nireq to slightly exceed the -nstreams value,
summed across all devices used.
NOTE: This demo is based on the callback functionality from the Inference Engine Python API. The selected approach makes the execution in multi-device mode optimal by preventing wait delays caused by the differences in device performance. However, the internal organization of the callback mechanism in Python API leads to a decrease in FPS. Please, keep this in mind and use the C++ version of this demo for performance-critical cases.
NOTE: If you provide a single image as an input, the demo processes and renders it quickly, then exits. To continuously visualize inference results on the screen, apply the
loopoption, which enforces processing a single image in a loop.
You can save processed results to a Motion JPEG AVI file or separate JPEG or PNG files using the -o option:
- To save processed results in an AVI file, specify the name of the output file with
aviextension, for example:-o output.avi. - To save processed results as images, specify the template name of the output image file with
jpgorpngextension, for example:-o output_%03d.jpg. The actual file names are constructed from the template at runtime by replacing regular expression%03dwith the frame number, resulting in the following:output_000.jpg,output_001.jpg, and so on. To avoid disk space overrun in case of continuous input stream, like camera, you can limit the amount of data stored in the output file(s) with thelimitoption. The default value is 1000. To change it, you can apply the-limit Noption, whereNis the number of frames to store.
NOTE: Windows* systems may not have the Motion JPEG codec installed by default. If this is the case, you can download OpenCV FFMPEG back end using the PowerShell script provided with the OpenVINO ™ install package and located at
<INSTALL_DIR>/opencv/ffmpeg-download.ps1. The script should be run with administrative privileges if OpenVINO ™ is installed in a system protected folder (this is a typical case). Alternatively, you can save results as images.
You can also run this demo with model served in OpenVINO Model Server. Refer to OVMSAdapter to learn about running demos with OVMS.
Exemplary command:
python3 object_detection_demo.py \
-i <path_to_video>/inputVideo.mp4 \
-m localhost:9000/models/object_detection \
-at ssd \
--labels <omz_dir>/data/dataset_classes/voc_20cl_bkgr.txt \
--adapter ovmsThe demo uses OpenCV to display the resulting frame with detections (rendered as bounding boxes and labels, if provided). The demo reports
- FPS: average rate of video frame processing (frames per second).
- Latency: average time required to process one frame (from reading the frame to displaying the results). You can use both of these metrics to measure application-level performance.