This is the demo application for Action Recognition algorithm, which classifies actions that are being performed on input video. The following pre-trained models are delivered with the product:
driver-action-recognition-adas-0002-decoder, which are models for driver monitoring scenario. They recognize actions like safe driving, talking on the phone and others
i3d-rgb-tf, which are general-purpose action recognition (400 actions) models for Kinetics-400 dataset.
The demo pipeline consists of several steps, namely
Render. Every step implements
PipelineStep interface by creating a class derived from
PipelineStep base class. See
steps.py for implementation details.
DataStepreads frames from the input video.
EncoderSteppreprocesses a frame and feeds it to the encoder model to produce a frame embedding. Simple averaging of encoder's outputs over a time window is applied.
DecoderStepfeeds embeddings produced by the
EncoderStepto the decoder model and produces predictions. For models that use
DummyDecoder, simple averaging of encoder's outputs over a time window is applied.
<ModelNameStep>does preprocessing and produces predictions.
RenderSteprenders prediction results.
Pipeline steps are composed in
AsyncPipeline. Every step can be run in separate thread by adding it to the pipeline with
parallel=True option. When two consequent steps occur in separate threads, they communicate via message queue (for example, deliver step result or stop signal).
To ensure maximum performance, Inference Engine models are wrapped in
AsyncWrapper that uses Inference Engine async API by scheduling infer requests in cyclical order (inference on every new input is started asynchronously, result of the longest working infer request is returned). You can change the value of
action_recognition_demo.py to find an optimal number of parallel working infer requests for your inference accelerators (Intel(R) Neural Compute Stick devices and GPUs benefit from higher number of infer requests).
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/action_recognition_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:
An example of using the Model Converter:
Running the application with the
-h option yields the following usage message:
Running the application with an empty list of options yields the usage message given above and an error message.
For example, to run the demo for in-cabin driver monitoring scenario, please provide a path to the encoder and decoder models, an input video and a file with label names, located in the demo folder,
>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
loop option, 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
aviextension, for example:
pngextension, 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_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 the
limitoption. The default value is 1000. To change it, you can apply the
-limit Noption, where
Nis 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.
The application uses OpenCV to display the real-time action recognition results and current inference performance (in FPS).