the yolov5 model performance runned by DJL is poor

[zhangkai253]

[frankfliu]

Are you using our model zoo model? Our model zoo won't be able to detect Chinese word. How you convert your model?
Can you share your java code that use DJL? How to do the pre/post processing?

Can you share your model and code for both java and Python?

[frankfliu]

Most likely this is related to YoloV5Translator, the image processing in YoloVeTranslator is different from your python code.

You might need to check if the Pipeline is matching what you are using in the python code.

Can you share your model and python code?

[frankfliu]

Can you share your model? Our built-in yolov5s model zoo model seems working fine.

[zhangkai253]

Yes, I tried the built-in model ever, it worked fine indeed. Can you give me your email adress? It seems that I cant upload the model here.

[frankfliu]

Can you upload to your git repo? the file might be too big for email.

[zhangkai253]

HI Frank, any idea?

[frankfliu]

Will take a look this weekend.

[frankfliu]

@zhangkai253

The current Yolov5Translator resize the image to 640x640, but detect.py pad the image to 640x640, this causing detection quality reduced.

You can manually pad your image to 640x640, and you will get the same result.

I created a PR to add a CenterFit image operation to solve this issue: #3425
Once this PR is merged you can use the follow code to get the result:

        Criteria<Image, DetectedObjects> criteria =
                Criteria.builder()
                        .setTypes(Image.class, DetectedObjects.class)
                        .optModelPath(Paths.get("best.pt"))
                        .optTranslatorFactory(new YoloV5TranslatorFactory())
                        .optArgument("height", 640)
                        .optArgument("width", 640)
                        .optArgument("centerFit", true)
                        .optArgument("nmsThreshold", 0.45)
                        .optArgument("threshold", 0.27)
                        .optArgument("synset", "pa,chong,wang,luo")
                        .optEngine("PyTorch")
                        .build();