c6480de736
* bind success * bind success fix * FDMat pybind, ResizeByShort pybind * FDMat pybind, ResizeByShort pybind, remove initialized_ * override BindProcessorManager::Run in python is available * PyProcessorManager done * vision_pybind fix * manager.py fix * add tutorials * remove Apply() bind * remove Apply() bind and fix * fix reviewed problem * fix reviewed problem * fix reviewed problem readme * fix reviewed problem readme etc * apply return outputs * nits * update readme * fix FDMatbatch * add op pybind: CenterCrop, Pad * add op overload for pass FDMatBatch --------- Co-authored-by: Wang Xinyu <shaywxy@gmail.com>
1.4 KiB
English | 中文
Vision Processor
Vision Processor is used to implement model preprocessing, postprocessing, etc. The following 3rd party vision libraries are integrated:
- OpenCV, general CPU image processing
- FlyCV, mainly optimized for ARM CPU
- CV-CUDA, for NVIDIA GPU
C++
TODO(guxukai)
Python
Python API, Currently supported operators are as follows:
- ResizeByShort
- NormalizeAndPermute
Users can implement a image processing modules by inheriting the PyProcessorManager class. The base class PyProcessorManager implements GPU memory management, CUDA stream management, etc. Users only need to implement the apply() function by calling vision processors in this library and implements processing logic. For specific implementation, please refer to the demo code.
Demo
Performance comparison between CV-CUDA and OpenCV:
CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz
GPU: T4
CUDA: 11.6
Processing logic: Resize -> NormalizeAndPermute
Warmup 100 rounds，tested 1000 rounds and get avg. latency.
| Input Image Shape | Target shape | Batch Size | OpenCV | CV-CUDA | Gain |
|---|---|---|---|---|---|
| 1920x1080 | 640x360 | 1 | 1.1572ms | 0.9067ms | 16.44% |
| 1280x720 | 640x360 | 1 | 2.7551ms | 0.5296ms | 80.78% |
| 360x240 | 640x360 | 1 | 3.3450ms | 0.2421ms | 92.76% |