insight-platform/Savant
Python Computer Vision & Video Analytics Framework With Batteries Included
What it solves
Savant simplifies the creation of high-performance, real-time multimedia AI applications on Nvidia hardware. It removes the need for low-level programming typically required by Nvidia DeepStream, allowing developers to build scalable, fault-tolerant inference pipelines for computer vision and video analytics that work across both edge devices (Jetson) and data centers.
How it works
Savant acts as a high-level abstraction layer built on top of Nvidia DeepStream. It uses a declarative approach where pipelines are defined as blocks via YAML configurations and implemented using Python functions. The framework provides a common protocol for video and metadata delivery, a Client SDK for integration, and a Development Server that allows for dynamic code reloading without restarting the pipeline.
Who it’s for
Computer vision and video analytics engineers who need to deploy production-ready applications on Nvidia equipment and want to reduce time-to-market by avoiding low-level SDK complexities.
Highlights
- Hardware Versatility: Supports Nvidia Jetson (Nano/NX/AGX) and data center GPUs (Turing, Ampere, Ada, Hopper, Blackwell).
- GPU-Accelerated Processing: Integrates PyTorch, CuPy, and OpenCV CUDA to perform inference and post-processing directly in GPU RAM, minimizing CPU-GPU data transfers.
- Extensive Connectivity: Includes a wide array of built-in source and sink adapters for RTSP, USB/CSI cameras, Kafka-Redis, and Amazon Kinesis Video Streams.
- Cloud-Native: Deploys via Docker and includes integrated monitoring and tracing using OpenTelemetry and Prometheus.
- Dynamic Management: Supports attaching/detaching data sources and updating pipeline parameters on the fly without reloading.