Hi I am working on a crowd counting project and I am using this repo. Congrats on the paper btw. I made some changes since I was using it I made some changes and thought it might be good to contribute. Here are the changes;
training/testing arguments can be loaded from a json file as stated in readme file. So you switch between configuration templates more easily.
added some documentation about it
removed crop-size argument from the cli since it is always overriden by dataset defaults.
also dataset defaults are in a seperate json file
minor refactoring to reduce the operation count for sinkhorns algorithm. Now self.cood * self.cood is stored in a constant variable so it won't always multiply again also commented out an assertion which is not usefull in our specialized case.
There are also some new things I am planning to do such as;
adding GCC dataset class (even though I am using one I might need to clean and create an issue for it first)
using it with another backbone
finding a way to vectorize therefore optimize the usage of sinkhorns algorithm implementation
maybe using a different backbone and sharing the results
Hi I am working on a crowd counting project and I am using this repo. Congrats on the paper btw. I made some changes since I was using it I made some changes and thought it might be good to contribute. Here are the changes;
There are also some new things I am planning to do such as;