OpenCV 4.9 Brings DNN Module Enhancements, Orbbec Gemini 2 Camera Support

OpenCV 4.9 released on Friday as the newest version of this widely-used, open-source computer vision (CV) library.

OpenCV developers are working toward OpenCV 5 with improved AMD CPU/GPU support, better Raspberry Pi support, enhanced Deep Neural Network (DNN) support, Android packaging, code refactoring, TrueType Font support, and other improvements.

OpenCV 4.9 meanwhile as an interim release is out for the New Year. OpenCV 4.9 fixes several rounding issues on ARM processors, adds detection and dispatching of some modern ARM NEON instructions (Neon FP16 & Neon BF16), and adds optimizations for LoongArch 128-bit vector detection/dispatching.

OpenCV 4.9's Deep Neural Network (DNN) module also adds experimental transformers support, various ONNX support improvements, an OpenVINO back-end for INT8 models, CANN back-end improvements, and the DNN module's layer normalization now supports OpenVINO / OpenCL / CUDA back-ends.

OpenCV 4.9 also brings several improvements to its object detection module, there is a new object tracking API (TrackerVit) that is vision transformer based, calibration module fixes, Python binding fixes, Android support improvements, experimental CUDA support as a first-class language in CMake, experimental support for the Apple VisionOS platform, and support for the Orbbec Gemini 2 and Gemini 2 XL cameras.

Orbbec's Gemini 2 is a 3D camera released this year that makes use of active stereo IR and a custom ASIC for high quality depth processing, IMU, and USB 3.0 Type-C connectivity and power. The Gemini 2 can handle up to 1280 x 800 @ 30 FPS while retailing for $299 USD.

OpenCV 4.9 downloads and more information on this updated computer vision library via OpenCV on GitHub.