caixiongjiang / FastSegFormer-pyqt

FastSegFormer-QTUI && Jetson Nano Deployment

FastSegFormer-pyqt

中文

Navel orange defect segmentation model video detection UI.

Update

• Create PyQT interface for navel orange defect segmentation. (May/10/2023)
• Produce 30 frames of navel orange assembly line simulation video. (May/13/2023)
• Support onnx format for video detection. (May/14/2023)
• Using multi-threaded processing, the main thread updates the UI and the sub-threads are used to process the video frames and improve the FPS to 48~60.(May/25/2023)
• Deployed on Jetson Nano(4G), an edge computing device with ONNXRuntime and TensorRT.(May/30/2023)
• Acceleration with DeepStream Framework on Jetson Nano(4G).(June/16/2023)

Demo

Navel orange simulation line detection video

Usage

• Environment Configuration：

$ conda activate 'your anaconda environment'
$ pip install -r requirements.txt 

• Run project:

python run_gui.py

• Jetson Nano Deployment: Usage

Testing performance comparison

All the following tests are not only network inference, but also include pre-processing and post-processing, and the processing of video frames may be different for different methods.

• System: Windows 10, CPU: Intel(R) Core(TM) i5-10500 CPU @ 3.10GHz GPU: NVIDIA GeForce RTX 3060(12G)

FastSegFormer-pyqt
Task	Video input	Inference input	Inference framework	GPU computing capability	Quantification	Video processing	Average FPS
Video Detection	$512\times 512$	$224\times 224$	PyTorch	12.74 TFLOPS	FP32	Single frame	32.62
			ONNXRuntime				32.64
			PyTorch				32.24
			ONNXRuntime				32.66
			PyTorch			Multi-thread	46.94
			ONNXRuntime				46.81

Conclusion：

1. In the mode of GPU for inference, there is almost no difference in inference time between PyTorch and ONNXRuntime.
2. There is almost no difference in GPU inference time using cards with the same FP32 and FP16 arithmetic power.
3. The main thread handles the input and output of the video, and the secondary thread handles the inference of the single-frame image, which has a great improvement on the detection performance of the video.

• System: Ubuntu 18.04 CPU: ARM Cortex-A57 @ 1.43GHz GPU: NVIDIA Maxwell @ 921MHz

Jetson-FastSegFormer
Task	Video/Stream input	Inference input	Inference framework	GPU computing capability	Quantification	Video processing	Average FPS
Video Detection	$512\times 512$	$224\times 224$	ONNXRuntime	0.4716 TFLOPS	FP16	Single frame	10
			TensorRT				15
			ONNXRuntime			Multi-thread	~
			TensorRT				~
			TensorRT			DeepStream	23
CSI Camera Detection	$1280\times 720$		ONNXRuntime			Single frame	8
			TensorRT				12
			ONNXRuntime			Multi-thread	~
			TensorRT				~
			TensorRT			DeepStream	20

~:Can't run dual-threaded acceleration on Jetson nano (4G) because of lack of memory.