qtimletflite

The QTI's qtimletflite GStreamer element exposes TensorFlow Lite (TFLite) capabilities to GStreamer. It can load and execute TFLite models. It supports preprocessing and postprocessing functionality. The preprocessing supports downscale, color convert, mean subtraction and padding. The postprocessing supports the most popular model types as classification, detection, and segmentation. The postprocessing result is attached as machine learning metadata (MLMeta) to GST buffer. The QTI's qtimletflite gstreamer element support two ways of configuration: config file in JSON format and GST properties. Each configuration parameter can be set in both ways, but GST properties has higher priority.

Pad templates

sink (input)

Presence – always

Direction – sink

source (output)

Presence – always

Direction – src

Properties

name

The name of the object.

qos

Handle Quality-of-Service events.

config

Path to JSON file. Eg.: /data/misc/camera/mle_tflite_config.json

model

Model file name. Path to model file. Eg.: /data/misc/camera/model.tflite. Filename extension should be .tflite

labels

Labels file name. Path to labels file. Eg.: /data/misc/camera/labels.txt. Filename extension should be .txt

input-format

Set inference input format. Supported formats: 0 - RGB; 1 - BGR; 2 - RGBFloat; 3 - BGRFloat.

postprocessing

Supported Postprocessing: classification; detection; segmentation; posenet

mean

Channel Mean Subtraction values ('<B, G, R>'). Eg. mean="<128.0,128.0,128.0>"

sigma

Channel divisor values ('<B, G, R>'). Eg. sigma="<128.0,128.0,128.0>"

preprocess-type

Controls pre-processing aspect ration(AR) maintenance.

Possible values: 0-kKeepARCrop, 1-kKeepARPad, 2-kDirectDownscale

confidence-threshold

Confidence Threshold value.

delegate

upported TFLite delegates: default - use CPU; dsp - use DSP

num-threads

Number of threads that are to be used for inference.