VPI-3.0.10 is the first production release of VPI-3.0 branch. It contains several new algorithms and bug fixes.
Documentation
Link to documentation for release VPI-3.0
Release Highlights
- New algorithms:
- Brute Force Matcher :
- Added implementation on CUDA backend.
- Significant improved performance on CPU backend by making use of all CPU cores available and further improvements in algorithm complexity when doing cross-check.
- Dense Optical Flow :
- Expose more parameters from SGM algorithm OFA implementation, to allow fine-tuning of output quality. See VPIOpticalFlowDenseSGMParams and related new APIs.
- Increased supported pyramid levels from 5 to 7, allowing detection of flow across larger areas.
- Python API:
- Binaries for python 3.9 and 3.10.
- Deployment of VPI's python source code so that it can be compiled for other Python versions.
- Supported x86 platforms:
- x86 Ubuntu 20.04 and 22.04
- Jetson boards:
- AGX Orin DevKit
- AGX Orin 32/64GB/JAOi
- Orin NX 16/8GB
- Orin Nano 8/4GB
- Orin Nano DevKit
Other Updates
- Colornames:
- Added PVA backend, which accepts RGB8p input image format.
- HOG:
- Up to 2x faster performance of CUDA backend.
- Added support for several RGB planar image formats like RGBA8p, RGB8p, BGRA8p and BGR8p. The following algorithms were updated:
- Gaussian Pyramid Generator :
Selected Bug Fixes
- Fixed segfault with python bindings when importing modules like
torch
after vpi's.
Removed Features
- Temporal Noise Reduction :
- TNRv2 on VIC backend is no longer supported.
- Stereo Disparity Estimator :
- Removed the following backends:
- CPU-only, PVA-only, PVA+NVENC+VIC.
- GUI demos have been removed.
- Xavier devices are no longer supported.
- Dropped support for CUDA-11.x. Only CUDA 12.x is supported now.
- Dropped support for Ubuntu 18.04
Known Issues
- Host images wrapped into VPIImages using vpiImageCreateWrapper might incur in a performance hit when using them with algorithms running on CUDA backend. User should avoid wrappers in this case, preferring to use VPIImages allocated with vpiImageCreate.
- Possible performance hit when using CUDA images wrapped into VPIImages using vpiImageCreateWrapper in algorithms running in PVA, VIC and/or OFA. User should avoid using wrappers in this case, preferring to use VPIImages allocated with vpiImageCreate.
- Harris Corner Detector result scores/positions might differ among backends.
- Stereo Disparity Estimator
- The confidence map generated by OFA+PVA+VIC backend might have some negligible differences with respect to other backends.
- Performance hit using Dense Optical Flow on python due re-creating the payload at every call.
- Per-algorithm performance tables weren't updated, they still refer to performance from vpi-2.0.
- Attention
- Algorithms running on PVA backend won't work inside a docker container. Submission calls will return error VPI_ERROR_INTERNAL .