The latest update from Savant, version 0.2.6, marks a significant milestone with the platform surpassing 300 stars on GitHub and an active community on Discord. This update, which was developed over a period of one and a half months, brings a host of new features and improvements that we will delve into.
For those interested in exploring the update, you can find it on GitHub at Savant’s Release Page.
🔗 Join our Discord and visit GitHub to learn more about Savant’s cutting-edge features and to stay updated with our latest advancements. Let’s shape the future of computer vision together.
Please go to our blog to read the original release announcement with more pictures, videos and links.
New Demos
Savant has expanded its demo offerings to a total of 22, with the latest release including:
- A demo for recognizing car license plates.
- A demo for detecting key points on a person’s body.
- An advanced traffic metering demo.
- A demo for counting objects within specific areas.
- An AnimeGAN demo for style transfer in animations.
- A demo for enhancing video resolution.
- A demo showcasing video pass-through with pipeline chaining.
Enhancements to Existing Demos
- The traffic meter demo now comes with a guide on model preparation.
- The Facial ReID demo has been improved with an index builder that is integrated with ClientSDK, ensuring image resolutions are automatically adjusted.
- The conditional video processing demo has been enhanced to demonstrate dynamic reconfiguration using Etcd.
Support for Jetson Orin Nano
The release has been thoroughly tested on the Nvidia Jetson Orin Nano, showing that it outperforms the previous generation Nvidia Jetson NX. Notably, the Jetson Orin Nano lacks the NVENC device, prompting the development of features to support devices without it, such as:
- Video pass-through processing that transfers video from input to output without the need for transcoding.
- A GPU-less Always-On RTSP adapter capable of handling RTSP transcoding without a GPU.
Other Noteworthy New Features
- The ClientSDK now includes asynchronous classes and supports PNG images.
- The Kafka/Redis adapter has been improved to work exclusively with Kafka and features data deduplication when used in conjunction with video pass-through.
- Nvidia’s tracker has been updated to reduce the occurrence of phantom objects, which can be configured using a global variable.
- Support for Etcd-based dynamic parameters has been added.
Documentation and Benchmarking
- The developer’s guide has been updated and bugs have been fixed.
- Documentation is now versioned with stable URLs for releases and development branches, with automatic updates for the development branch.
- Benchmarking has been standardized for consistency, with new benchmarks excluding frame drawing for more accurate results.
Looking Ahead
For the upcoming 0.2.7 release, Savant plans to:
- Introduce a persistent buffer to manage bursts and outages without relying on Kafka.
- Add a new inference block for low-level inference using TensorRT.
- Showcase the integration of PyTorch and CuPY within the Savant framework.
- Stay updated with Savant by subscribing to their updates and joining the Discord community for support.
For further reading on why Savant might be the right choice for your computer vision projects, consider their article “Ten reasons to consider Savant for your computer vision project.”