ncnn
ncnn copied to clipboard
Tencent
Wrong result for onnx-ncnn yolov5s_6.0
Problem: incorrect detection
Windows platform. Image with wrong detection for example: 
model | 模型 | モデル
- original model: https://github.com/ultralytics/yolov5
how to reproduce | 复现步骤 | 再現方法
- Install yolov5 steps from Quick Start Examples.
- Verify:
python detect.py --source data/images --weights yolov5s.pt --conf 0.25. - Export for solve problems 'Unsupported slice step !':
python export.py --include onnx --weight yolov5s.pt --img 640 --train --simplify. According to this. - Conversion using out-of-the-box web model conversion (https://convertmodel.com/#outputFormat=ncnn): check both 'Optimize', check 'Generate ft16 model'.
- Comparing my .param file with the param file example (https://github.com/nihui/ncnn-assets/blob/master/models/yolov5s_6.0.param). The layers are the same, only the numbers differ. Size the .bin file is the same as in the example
- Use ncnn tutorial (https://github.com/Tencent/ncnn/blob/master/examples/yolov5.cpp) for yolov5s_6.0. Rewrite
ex.extract()according to my .param file numbers. - ...And I get a completely incorrect detection result :(
@zhiliu6, could you tell me how you converted the model? it seemed to me that I was doing similar steps, but this is not the case :(
I have similar problems,run on android,params yololv5 6.0 by pnnx
Running into a similar issue. Have you had any luck resolving this @xbc922 @Draft00 @zhiliu6?
Running into a similar issue. Have you had any luck resolving this @xbc922 @Draft00 @zhiliu6?
@lizozom, Nope :(. I had to use yolov4.
@Draft00 my intuition tells me that the processing of images is different between the python and JavaScript versions.
I'm not an ML pro, so I'm having a hard time analysing this myself.
