Video Processing Workflow

In Savant every frame passes certain processing stages which you have to understand. These stages are inspired by DeepStream’s internals and there is no simple way to hack them in a different way.

Those stages are (see the image below):

  • decoding (3);

  • scaling to a common resolution (mandatory) (4);

  • adding commonly-specified paddings (optional) (4);

  • multiplexing (4-5);

  • processing (6);

  • drawing (optional) (6);

  • de-multiplexing (7, 8);

  • encoding (9).

Discuss them in details.

Decoding

Nvidia supports very fast hardware-accelerated decoding for several video/image codecs. The dedicated NVDEC device performs the task at speed higher than 1000 FPS for FullHD. You must consider using Savant with those codecs which have hardware support. We currently support several source formats for the frames. The framework automatically understands the formats; you may simultaneously feed the module with streams in different codecs.

Raw RGBA is a slow representation as it requires extensive data transfers over PCI-E and network, leading to decreased performance.

Hardware-accelerated with NVDEC, preferred to be used:

  • H264: default;

  • HEVC/H265: preferred (performance, bandwidth);

  • MJPEG, JPEG: when image streams or USB/CSI-cams are used.

Software-decoded (not recommended to use):

  • PNG: fallback for compatibility purposes.

Scaling to a Common Resolution

The pipeline will scale all streams to the configured resolution.

This is a crucial topic. The pipeline is always configured to run on common resolution. It means that every stream handled by a certain pipeline instance is always scaled to the common resolution configured for the pipeline instance, no matter what its input resolution was.

If you need different streams to be handled on different resolutions, you have to launch several pipelines configuring each pipeline to use a resolution acceptable for streams processed by that pipeline.

# base module parameters
parameters:
  # pipeline processing frame parameters
  frame:
    width: 1280
    height: 720

Let us consider the following examples:

Case 1: No Output Footage

You have 10 cams of FullHD and 15 cams of HD resolution. You don’t need the outgoing video at all, all your models are fine to work with HD resolution.

Solution: configure the pipeline to use HD resolution, send all streams to a single pipeline.

Case 2: Low-Res Output Footage

You have 10 cams of FullHD and 15 cams of HD resolution. You need the outgoing video and HD resolition is acceptable, all your models are fine to work with HD resolution.

Solution: configure the pipeline to use HD resolution, send all streams to a single pipeline.

Case 3: Hi-Res Output Footage

You have 10 cams of FullHD and 15 cams of HD resolution. You need the outgoing video in the same resolution as incoming.

Solution: configure two pipelines - the first to use Full-HD resolution, the second to use HD resolution. Point Full-HD cams to the Full-HD pipeline, HD cams to the HD pipeline.

Paddings

Adding paddings is useful if you need spare space for utility purposes. E.g. you may use paddings to preprocess the image before passing it to the model. Another way to use paddings is to display utility content.

The paddings can either be preserved or removed at the output.

# base module parameters
parameters:
  # pipeline processing frame parameters
  frame:
    width: 1280
    height: 720
    # Add paddings to the frame before processing
    padding:
      # Paddings are kept on the output frame
      keep: true
      left: 0
      right: 1280
      top: 0
      bottom: 0

Note

If you specify parameters.frame.padding.keep == false, the paddings are removed before frame encoding. The geometry for all objects are recalculated to conform new geometry.

Geometry Base

The geometry_base parameter specifies the value by which any geometry dimension of the frame (width, height, margin size) must be evenly divided. The default value is 8.

When the developer specifies the frame dimensions do not fit the geometry_base, the pipeline will stop with an error. Thus, when defining frame.width, frame.height, and frame.padding.* every of them must be divisible by geometry_base. The parameter is introduced to overcome unexpected behavior due to platform-specific hardware limitations when a non-standard resolution is used during image processing and encoding.

Tip

We do not recommend setting geometry_base parameter to the values other than 8 or 4.

# base module parameters
parameters:
  # pipeline processing frame parameters
  frame:
    width: 1280
    height: 720
    # Base value for frame parameters. All frame parameters must be divisible by this value.
    # Default is 8.
    geometry_base: 8

Multiplexing

All streams processed by a single module instance are grouped into batches before processing. Batch is a concept used to optimize the computations on Nvidia hardware. Savant is implemented to hide batching: developers typically work with a single frame, not a batch of frames.

# base module parameters
parameters:
  ...
  batch_size: 1

Typically you may set batch_size equal to the maximum expected number of simultaneously processed streams. Find out more on Batching in the advanced topics.

Processing

At this step, the pipeline is implemented: model inference, tracking, and python functions work here. We will discuss the processing in detail in further sections. You cannot modify the frame resolution at this step. You can modify the frame content.

Drawing

Certain objects during the processing step can receive labels specifying that they must be drawn on the frame (e.g. identifiers, class names, boxes, etc). Drawing is an optional step which can be performed.

# base module parameters
parameters:
  ...
  draw_func: {}

The draw function may be overriden by the developer if the stock version cannot draw the information required:

parameters:
  ...
  draw_func:
    module: samples.peoplenet_detector.overlay
    class_name: Overlay
    kwargs:
      person_with_face_bbox_color: [0, 1, 0]
      person_no_face_bbox_color: [1, 0, 0]
      person_label_bg_color: [1, 0.9, 0.85]
      person_label_font_color: [0, 0, 0]
      bbox_border_width: 3
      overlay_height: 180
      logo_height: 120
      sprite_height: 120
      counters_height: 85
      counters_font_thickness: 5

Note

To disable draw_func functionality, remove parameters.draw_func from the manifest completely.

Conditional Drawing

Savant supports a conditional drawing feature. It enables defining a special condition based on a frame tag which enables drawing. The motivation behind the feature is efficiency: often, you don’t need to produce footage for all streams but only for certain streams under investigation. So you may implement a pyfunc which creates a tag for those streams.

To configure conditional drawing, add a subsection to draw_func as follows:

draw_func:
  condition:
    tag: <tagname, e.g. draw>

An example of conditional drawing can be found in a dedicated Savant sample.

De-Multiplexing

This step is automatically performed by the framework to turn batches into individual streams before passing the frames to stream encoders.

Encoding

If the output_frame section is omitted, video frames will not be sent to sinks at all.

The framework supports several encoding schemes:

  • RAW RGBA (not optimal, as it requires large transfers over PCI-E);

  • JPEG (hardware nvjpegenc, software jpegenc);

  • PNG (software pngenc);

  • H264 (hardware nvv4l2h264enc, software x264enc);

  • HEVC/H265 (hardware nvv4l2h265enc);

  • COPY (pass-through mode, the module sends frames from the source to the sink as is).

Note

Hardware encoder for JPEG is available only on Nvidia Jetson. On dGPU JPEG encoder is CUDA-assisted when supported by the hardware.

We highly advise using hardware assisted codecs. The only caveat is to steer clear from GeForce GPUs in production as they have a limitation constraining simultaneous encoding to 3 streams. In case you are using GeForce, choose RAW RGBA.

parameters:
  output_frame:
    codec: h264

You can choose hardware or software encoder by setting encoder parameter to nvenc or software respectively:

parameters:
  output_frame:
    codec: h264
    encoder: nvenc

When encoder parameter is specified and the framework doesn’t find a suitable encoder, it will end with an error. When encoder parameter is omitted, the framework will try to use hardware encoder. When it fails, it will fall back to software encoder.

Every codec has its own configuration parameters related to a corresponding GStreamer plugin. Those parameters are defined in output_frame.encoder_params:

parameters:
  output_frame:
    codec: h264
    encoder_params:
      bitrate: 4000000
      iframeinterval: 10
      profile: High

Tip

Find out more on the software H264 encoder on Medium.

Encoder Properties

Hardware H264 Encoder (NVENC)

  1. bitrate

    Sets the bitrate for the v4l2 encoder. Allowed range: 0 - 4294967295. The Default value is 4000000.

  2. control-rate

    Sets the control rate for the v4l2 encoder. The default value is 1.

    Options are:

    • 0 or variable_bitrate;

    • 1 or constant_bitrate;

  3. extended-colorformat

    Sets Extended ColorFormat pixel values 0 to 255 in VUI info. The default value is false.

  4. force-idr

    Forces an IDR frame. The default value is false.

  5. force-intra

    Forces an INTRA frame. The default value is false.

  6. iframeinterval

    Encoding Intra Frame occurrence frequency. Range: 0 - 4294967295. The default value is 30.

  7. preset-id

    Sets CUVID Preset ID for the encoder. Range: 1 - 7. The default value is 1.

  8. profile

    Sets the profile for the v4l2 encoder. The default value is 0 (Baseline).

    Options are:

    • 0: Baseline

    • 2: Main

    • 4: High

    • 7: High444

  9. tuning-info-id

    Tuning Info Preset for the encoder. The default value is 2.

    Options are:

    • 1: HighQualityPreset

    • 2: LowLatencyPreset

    • 3: UltraLowLatencyPreset

    • 4: LosslessPreset

Software H264 Encoder

  1. bitrate

    Bitrate in kbit/sec. Range: 1 - 2048000. The default value is 2048.

  2. key-int-max

    Maximum distance between two key-frames (0 for automatic). Range: 0 - 2147483647. The default value is 0.

  3. pass

    Encoding pass/type. The default value is 0 (cbr)

    Options are:

    • 0 or cbr: Constant Bitrate Encoding

    • 4 or quant: Constant Quantizer

    • 5 or qual: Constant Quality

    • 17 or pass1: VBR Encoding - Pass 1

    • 18 or pass2: VBR Encoding - Pass 2

    • 19 or pass3: VBR Encoding - Pass 3

  4. speed-preset

    Preset name for speed/quality tradeoff options (can affect decode compatibility - impose restrictions separately for your target decoder). The default value is 6 (or medium).

    Options:

    • 1 or ultrafast;

    • 2 or superfast;

    • 3 or veryfast;

    • 4 or faster;

    • 5 or fast;

    • 6 or medium;

    • 7 or slow;

    • 8 or slower;

    • 9 or veryslow;

    • 10 or placebo;

  5. tune

    Preset name for non-psychovisual tuning options. The default value is 0x00000000 or none.

    Options:

    • 0x00000000 or none

    • 0x00000001 or stillimage: Still image

    • 0x00000002 or fastdecode: Fast decode

    • 0x00000004 or zerolatency: Zero latency

Note

For this encoder a profile can be specified in output_frame.profile parameter (one of baseline, main, high). The default value is baseline.

Hardware HEVC Codec (NVENC)

  1. bitrate

    Sets the bitrate for the v4l2 encoder. Range: 0 - 4294967295. The default value is 4000000.

  2. control-rate

    Sets the control rate for the v4l2 encoder. The default value is 1 or constant_bitrate.

    Options are:

    • 0 or variable_bitrate;

    • 1 or constant_bitrate;

  3. extended-colorformat

    Sets extended color format pixel values 0 to 255 in VUI info. The default value is false.

  4. force-idr

    Forces an IDR frame. The default value is false.

  5. force-intra

    Forces an INTRA frame. The default value is false.

  6. iframeinterval

    Encoding Intra Frame occurrence frequency. Range: 0 - 4294967295. The default value is 30.

  7. preset-id

    Sets CUVID Preset ID for Encoder. Range: 1 - 7. The default value is 1.

  8. profile

    Sets the profile for the v4l2 encoder. The default value is 0 or Main.

    Options are:

    • 0 or Main

    • 1 `` or  ``Main10

  9. tuning-info-id

    Tuning Info Preset for the encoder. The default value is 2 or LowLatencyPreset.

    Options are:

    • 1 or HighQualityPreset

    • 2 or LowLatencyPreset

    • 3 or UltraLowLatencyPreset

    • 4 or LosslessPreset

JPEG Codec

  1. idct-method

    The IDCT algorithm to use. The default value is 1 or ifast.

    Options are:

    • 0 or islow: slow but accurate integer algorithm

    • 1 or ifast: faster, less accurate integer method

    • 2 or float: floating-point, accurate, fast on fast HW

  2. quality

    Quality of encoding. Range: 0 - 100. The default value is 85.

PNG Сodec

  1. compression-level

    PNG compression level. Range: 0 - 9. The default value is 6.

Example:

parameters:
  output_frame:
    codec: h264
    encoder_params:
      bitrate: 4000000
      profile: 4
parameters:
  output_frame:
    codec: jpeg
    encoder_params:
      quality: 90

To list all available properties run gst-inspect-1.0 <encoder-name>. E.g. gst-inspect-1.0 nvv4l2h264enc.

Conditional Encoding

Savant 0.2.4 introduced a conditional encoding feature. It enables defining a special condition based on a frame tag, enabling encoding only certain streams. The motivation behind the feature is efficiency: often, you don’t need to produce a resulting video for all streams but only for certain streams under investigation. So you may implement a pyfunc which creates a tag for those streams.

To configure conditional encoding, add a subsection to output_frame as follows:

output_frame:
  codec: h264
  encoder_params:
    iframeinterval: 25
  condition:
    tag: <tagname, e.g. encode>

An example of conditional drawing can be found in a dedicated Savant sample.

Pass-through mode

Pass-through mode is a special mode when the module doesn’t encode the frame but passes it to the sink as is. This mode is useful when your module doesn’t modify the frame but only adds some metadata to it.

To configure pass-through mode, set output_frame.codec to copy:

output_frame:
  codec: copy

Note

Drawing on frames is not ignored in pass-through mode but the frame modifications exist only in the pipeline and are not propagated through the sinks.