OpenCV is Intel's open source computer vision library, which implements many common algorithms in image processing and computer vision. The main updates of OpenCV 4.9.0 version are as follows:
-
Core Module:
- #23965 added
cv::broadcast
- #23965 added
- DNN module patches:
- Experimental transformers support
- #24476 ONNX Attention layer support
- #24037 ONNX Einsum layer support
- #23987 OpenVINO backend for INT8 models
- #24092 ONNX Gather Elements layer
- #24378 ONNX InstanceNorm layer
- #23897 #24694 #24509 New fastGEMM implementation and multiple layers on top of it
- #23654 Winograd fp16 optimization on ARM
- Testing and multiple fixes for Yolo series model support
- New layer support and bug fixes in CUDA backend: GEMM, Gelu, Add
- #24462 CANN backend: bug fix, support HardSwish, LayerNormalization and InstanceNormalization
- #24552 LayerNormalization: Support OpenVINO, OpenCL and CUDA backends
-
G-API module:
- TBD
-
Objdetect module:
-
Video:
-
Calibration module:
- Multiple fixes and improvements to chess board calibration rig detector.
- #23025 If the calibration system constraints are insufficient, calibrateCamera will throw an exception.
- #24482 Fixed bug in USAC’s findEssentialMat
- #24527 Fixed
cv::cornerSubPix
out-of-image access in - #23607 Fixed crash in ap3p
- #24035 Fix stereoRectify image boundaries
- #24211 Fixed “use after free” issue in Essential_solver.cpp
-
Python bindings:
-
Android:
- New Android Archive Package (AAR) released via Maven Central
- New Android examples: QR code detector, video IO examples. DNN and face detector examples updated.
- Switched to Gradle 7.6.3, the modern Android tool.
-
Platform support:
- #23021 Experimentally support CUDA as a first class language in CMake
-
other:
- TBD
- OpenCV Summer of Code: HAN Liutong’s semi-automatic refactoring across multiple pull requests makes CPU-optimized code compatible with SIMD of variable vector length (RISC-V RVV)
See ChangeLog for more details .