Savant 0.2.6 was released on November 8, 2023. The release includes multiple bug fixes, seven new demos, and many other enhancements, including documentation, benchmarking, and Jetson Orin Nano support.
Savant crossed the 300-star band on GitHub, and Discord is now active. The work on the release took 1.5 months. In the following sections, we will cover essential parts of the release in detail.
GitHub: https://github.com/insight-platform/Savant/releases/tag/v0.2.6
New Demos
In total, we now provide 22 demos and samples. The most significant samples are collected on the Savant website.
The release introduces new demos:
- Car license plate recognition demo (YOLOV8N, NGC LPD, NGC LPR);
- Person body keypoint detection demo (YOLOV8N-POSE);
- Advanced traffic meter demo (YOLOV8M);
- Object counting by areas demo (PeopleNet);
- AnimeGAN demo (AnimeGANv2, Miyazaki style);
- Video super-resolution demo (NinaSR x3);
- Video pass-through demo (a multi-stage demo showing how to use a brand new video pass-through feature and pipeline chaining).
Improved Demos
- Traffic meter demo – we added instructions on how to prepare the model;
- Facial ReID demo – ReID index builder is now implemented with ClientSDK and automatically fixes resolutions of database images;
- The conditional video processing demo improved to show how Etcd can be used to reconfigure processing dynamically (read more).
Jetson Orin Nano Top Tier Support
We finally received the device in our lab, and this release is thoroughly tested to work correctly with it. All benchmarks show that Nvidia Jetson Orin Nano is more capable than the previous generation Nvidia Jetson NX.
Jetson Orin Nano does not contain the NVENC device onboard, so we implemented a couple of features to support such devices (and A100/V100/H100/A30):
- Video pass-through processing – the pipeline copies video from input to output without transcoding (feature description);
- GPU-less Always-On RTSP adapter – the adapter now optionally can handle RTSP transcoding without GPU (feature description);
New Significant Features
- ClientSDK now contains asynchronous classes;
- ClientSDK now supports PNG images;
- The Kafka/Redis adapter now can work with Kafka only and supports data deduplication in Kafka/Redis mode when video pass-through is used;
- Nvidia’s tracker was modified to avoid creating phantom objects by default, configurable with a global variable;
- video pass-through processing – the pipeline copies video from input to output without transcoding (feature description);
- GPU-less Always-On RTSP adapter – the adapter now optionally can handle RTSP transcoding without GPU (feature description);
- Etcd-based dynamic parameters support;
Documentation
- We updated the developer’s guide and fixed bugs;
- Now, the documentation is versioned: releases and development branches have stable URLs, and documentation for the development branch is updated automatically;
Benchmarking
We unified benchmarking to get more consistent results. All the benchmarks now do not contain drawing on frames, and here are the fresh numbers:
| Benchmark | A4000 FPS | Jetson Xavier NX FPS | Jetson Orin Nano FPS |
| Age/Gender Prediction | 369 | 48 | 72 |
| Facial ReID | 225 | 36 | 50 |
| Intersection Traffic Meter | 261 | 32 | 46 |
| License Plate Recognition | 271 | 35 | 43 |
| Car Classification | 464 | 64 | 132 |
| MOG2 Background Removal | 743 | 102 | 137 |
| PeopleNet Detector | 414 | 77 | 131 |
| Traffic Meter | 257 | 54 | 101 |
| YOLOV8 Segmentation | 81 | 24 | 38 |
Future Release
In the future 0.2.7, we plan:
- Introduce a new building block: a persistent buffer helping to fight bursts and outages without the need for Kafka;
- Introduce a new inference block: low-level inference based on TensorRT;
- Demonstrate how to use PyTorch in Savant;
- Demonstrate how to use CuPY in Savant.
Don’t forget to subscribe to our X to receive updates on Savant. Also, we have Discord, where we help users.
To know more about Savant, read our article: Ten reasons to consider Savant for your computer vision project.