Different outputs generated by SiamMaskCpp and SiamMask python version

[atishay14]

Hi,

I have been trying to run the code in this repository but from my observation the accuracy of the CPP version seems to be worse than the python version for the same input. I suspected two things:

1. Model Conversion from PyTorch to Torch Script:

I tried running the convert tool on the refined models in the SiamMask repository at the following link:
https://github.com/jiwoong-choi/SiamMask#converting-siammask-model-with-the-refine-module-to-torch-scripts

However, I seem to get the dumped module files to be different in size than the ones provided in the repository (https://github.com/nearthlab/SiamMaskCpp/releases/download/v1.0/SiamMask_DAVIS.tar.gz).

The output is as follows:

feature_downsampler.pt - 1039 KB
feature_extractor.pt - 54,097 KB
mask_conv_kernel.pt - 2,319 KB
mask_conv_search.pt - 2,319 KB
mask_depthwise_conv.pt - 3 KB
refine_model.pt - 10,468 KB
rpn_model.pt - 9,812 KB

Downloaded torch script modules:

feature_downsampler.pt - 1096 KB
feature_extractor.pt - 57,056 KB
mask_conv_kernel.pt - 2,360 KB
mask_conv_search.pt - 2,360 KB
mask_depthwise_conv.pt - 3 KB
refine_model.pt - 11,210 KB
rpn_model.pt - 10,428 KB

Also the fps that I get when I use these models is one-fourth of the fps that I get when I use the models referred for download in the repository. Am I doing something wrong?

2. C++ code Implementation:

Is C++ code implementation different from Python code implementation?