Tensorrt_yolo3 works only with pre-built fp32.engine file

[sergiev]

With fp32.engine file downloaded from your cloud storage everything is working fine, but when I remove it and let caffe build the engine on its own, output becomes incredibly messy:


Currently I'm working on integration of model trained on custom dataset, could you please explain how to provide the compatibility?

[yukkysaito]

@k0suke-murakami

[k0suke-murakami]

Inside implementation is from here.
https://github.com/lewes6369/TensorRT-Yolov3

Please check the description over there.

[sergiev]

@k0suke-murakami could you please enlight me on the source of the .caffemodel you use to build .engine file. It seems like converted from darknet .weights file.
So,

1. What kind of converter was used?
2. The original .weights - is it the official yolov3-416 model from pjreddie's site or something different?