Python framework Savant 0.2.5 has been released, simplifying the use of Nvidia deepstream for machine learning problem solving. Savant handles the complex work with Gstreamer or FFMPEG, allowing users to focus on building optimized output conveyors using declarative syntax (YAML) and Python functions. With Savant, users can create conveyors that work on both accelerators in the data center and EDGE devices, allowing for easy processing of multiple video streams and quick creation of video analytics conveyors for working applications using NVIDIA Tensor. The project code is distributed under the Apache 2.0 license.[1]
The main innovations include improvements to Nvidia Deepstream 6.3, developer tools, adapters, and the addition of new examples and demonstrations. The improvements in developer tools include Opentelemetry support for journaling frame processing, hot reboot of changed Python code in development mode, and synchronous and asynchronous Client SDK for testing and accessing adapters. The loggia has also been improved with a color release.[1]
The improvements in adapters include a data processing adapter using Kafka/Redis for high load conveyors, a performance testing adapter (Multi-Stream Source Test Adapter), a Python SDK for developing custom adapters, simultaneous processing of multiple flows in the RTSP video broadcasting adapter, and the addition of HEVC coding to the adapter for accessing Gige Vision cameras.[1]
New examples and demonstrations have also been added, including real-time instance segmentation based on Yolov8m-Eg, re-identification by face using Yolov8-Face, Adaface, and HnSwlib, forecasting gender and age using Yolov8-Face and Mobilenet V2, and improved Traffic Meter example with a choice of three detectors – PEOPLENET, YOLOV8S, and Yolov8m.[1]
There are also auxiliary examples available, such as using Opentelemetry, using Client SDK, a simple conveyor for determining RTSP camera compatibility with Savant, and an example of using kafka/redis adapters.