News, updates, and handy materials related to the Savant ecosystem
A SegFormer-B2 model trained on nothing but Blender renders, with zero hand-labelled images: 0.9950 IoU and 0.9973 pixel accuracy on held-out synthetic data, and 97-99% on real photographs. The label...
Read more →
Light blur improves tracking accuracy for 17 of 23 trackers, and a streak encodes velocity you can measure from one frame. Then the deployment arithmetic: a 4K frame is 32...
Read more →
Twenty years of association research, and on small objects the winning configuration is a 2016 Kalman filter with the IoU replaced. Re-identification is not weak at six pixels, it is undefined — and on the hard single-object benchmark a zero-shot segmentation model beats a 1.2-billion-parameter supervised tracker.
Read more →
On AI-TOD, changing the label-assignment metric on an unchanged Faster R-CNN is worth +10.0 mAP. Upgrading the architecture to a far heavier model is worth +3.7. The ordering of that comparison is the whole finding — here is the assignment, tiling and temporal work behind it, and where the detector...
Read more →
A default checkpoint with default settings is close to a worst case for small objects. Eleven specific defaults work against you, most of them silently — including one version-specific assignment hole that gives 75% of four-pixel targets zero positive labels. Plus the tracker configs, which are worse.
Read more →
Five independent failures compound when the target is six pixels wide, and none of them is architectural. The IoU cliff, stride mismatch, resolution destroyed before inference, missing signal, and annotation noise as the accuracy ceiling — plus the sixteen fields where this is the whole job.
Read more →
A new Savant sample detects when a camera has been moved, blocked, or its scene has changed — by comparing a ReID embedding of a region of interest against the same region a few seconds earlier. Dynamic per-source ROIs over Etcd, x86 and Jetson bundles, and a single embedding per...
Read more →
The memory-safety pitch for Rust in computer vision misses the point. The real case is narrower and stronger: Rust governs the authored layer above the C/C++/CUDA substrate, and its compile-forced total handling is the most direct answer to the domain's central problem — the failure is always in the case...
Read more →
Savant output converters can now receive the frame's metadata — source_id, PTS, objects, and more — unlocking per-source and per-frame post-processing without breaking existing converters.
Read more →
Savant's new pygroup element runs several sequential PyFuncs inside a single GStreamer element — cutting inter-element overhead while keeping a separate telemetry span for every stage.
Read more →
A fact-checked, license-by-license guide to shipping YOLO in a commercial product: permissive vs. copyleft, what copyright actually protects, which YOLO flavors are safe (YOLOX, RT-DETR, DAMO-YOLO, PP-YOLOE) versus traps (Ultralytics AGPL, GPL v6/v7/v9, YOLO-NAS weights), the COCO/ImageNet data risk, and how to build a detector you fully own.
Read more →
A decision guide for NVIDIA Jetson Orin — when on-device neural inference makes Orin essential, how to move from dev kit to production module, the NVDEC/NVENC/NVJPEG media-engine trap, the CUDA-to-edge workflow, OS and lifecycle planning, and how Orin compares to Coral, Hailo, Rockchip, and Thor.
Read more →
A use-case-first guide to picking a computer vision camera that actually works in production — covering the five deployment classes, the three tradeoffs on every spec sheet, mounting rules that beat the spec sheet, latency, and real LPR examples.
Read more →
Savant now supports direct CSI camera capture on NVIDIA Jetson through a new nvarguscamerasrc_bin pipeline source element. Frames land in GPU memory with zero encoding or network overhead — the shortest path from sensor to inference for robotics and embedded vision.
Read more →
Hardware H.264 encoding works on the host but fails inside Docker on Jetson Orin NX with JetPack 6.2.2. The root cause turns out to be a missing two-megabyte package — and finding it required strace, DRM ioctls, and reverse-engineering NVIDIA's closed-source libraries.
Read more →
DeepStream is the most optimized SDK for video analytics on NVIDIA hardware. Savant is built on top of it. This article explains the relationship between the two, what each brings to the table, and which to reach for depending on your project's requirements.
Read more →
PyTorch and Savant are both Python-first and CUDA-capable, yet they solve fundamentally different problems. This article explains the design philosophies, target audiences, and use cases of each — and when it makes sense to use them together.
Read more →
A step-by-step guide to building a fan-out/fan-in video analytics pipeline that distributes a stream across multiple parallel modules, runs independent inference, and merges results back into a single frame using Savant and savant-rs.
Read more →
Savant ecosystem gains Meta-Merge, a new service that completes the Fan-Out/Fan-In pattern by merging metadata from parallel inference branches back into a single unified frame.
Read more →
A step-by-step guide to exporting Ultralytics models to the ONNX format for use in Savant's TensorRT-based inference pipeline.
Read more →
Savant 0.6.0 brings Blackwell GPU support through custom-built DeepStream 7.1 with TensorRT 10.9, plus Python 3.12 for improved CPU-bound performance.
Read more →
An in-depth landscape overview of current and legacy tech stacks for computer vision applications, covering frameworks from OpenCV to DeepStream to Savant.
Read more →
Savant ecosystem expands with Router, a new microservice for simplifying and optimizing the routing of video streams across complex CV pipelines.
Read more →
Retina RTSP Service enables seamless integration of real-time video streaming into Savant-powered AI video pipelines with low latency and high performance.
Read more →
Savant 0.5.1 brings DeepStream 7.0 support, 34 real-world samples, Key-Value API, original resolution processing, and many more enhancements.
Read more →
We explain the advantages of USB/CSI cameras over RTSP cameras for projects working on Nvidia Jetson Orin hardware.
Read more →
A deep dive into camera properties that matter for computer vision: sensor quality, lens selection, chip processing, and software encoding.
Read more →
A minor bugfix update to Savant 0.4.x, fixing issues with JSON/image/video sink adapters.
Read more →
A minor bugfix update to Savant 0.4.x for JSON/image/video sink adapters.
Read more →
Savant 0.4.1 continues the 0.4.x release cycle with new features, bug fixes, and a new watchdog service for pipeline health monitoring.
Read more →
Learn how to record and replay video streaming traffic for reproducible testing, troubleshooting, and code improvement in Savant pipelines.
Read more →
Savant now supports hardware-accelerated video transcoding and composition, enabling custom dashboards, resolution changes, and video walls.
Read more →
Savant 0.4.0 focuses on system usability, interfacing, advanced computer vision, and video analytics on top of DeepStream 6.4.
Read more →
New adapters integrate Amazon Kinesis Video Streams with Savant, combining video-optimized streaming and storage in the cloud.
Read more →
Learn how to dynamically attach and detach video sources to a running Savant pipeline using Docker, including Unix domain sockets and TCP/IP.
Read more →
Savant 0.3.11 adds MJPEG USB camera support -- a popular format for high-resolution, high-FPS streaming and stereo cameras.
Read more →
How to simulate an MJPEG USB camera in Linux using FFmpeg and a V4L2 loopback device for development and testing.
Read more →
Benchmark results demonstrating Savant's performance advantage over PyTorch for video inference, including hardware-accelerated decoding.
Read more →
Learn how to implement keypoint detection for pose estimation, facial recognition alignment, and action recognition using Savant.
Read more →
Using Nvidia PeopleNet and Savant to accurately count and monitor people in user-defined polygonal areas for various applications.
Read more →
Building a license plate detection and recognition pipeline using YOLOv8 and Nvidia NGC models with the Savant framework.
Read more →
Media Pass-Through allows copying the incoming video stream to output without transcoding, optimizing pipelines that don't need draw functionality.
Read more →
Why you may find it beneficial to use Savant instead of DeepStream, OpenVINO, PyTorch, or OpenCV in your next computer vision project.
Read more →
How to build scalable computer vision applications using Kafka and KeyDB for handling real-time and delayed video processing at scale.
Read more →
Exploring the challenges of real-time video analytics and practical approaches to overcome them using modern frameworks and hardware acceleration.
Read more →
Build a pipeline that calculates vehicles passing through metering zones and sends statistics to Grafana for visualization.
Read more →
Introducing RocksQ, a high-performance persistent queue built on RocksDB for buffering video frames and metadata in Savant.
Read more →
How to dynamically reconfigure pipeline parameters at runtime using Etcd for adaptive computer vision applications.
Read more →
Savant 0.2.5 brings significant changes, new features, and improvements to developer experience and deployment.
Read more →
How to use OpenTelemetry to trace, log, and collect metrics for deep learning computer vision pipelines in Savant.
Read more →
A high-performance facial re-identification pipeline using Savant for doorbell security and video content annotation systems.
Read more →
Running real-time instance segmentation with the YOLOv8 family on Savant, enabling efficient object segmentation on edge and data center hardware.
Read more →
How Savant handles video encoding on hardware that lacks NVENC, such as Jetson Orin Nano, Tesla A100, and Tesla H100.
Read more →
Element Groups allow conditionally loading pipeline elements during initialization based on environment variables and expressions.
Read more →