How to use my own Rugby Dataset ? What is the Input requirement and How to train and generate my own dataset mat files?
shaziagul-rgb opened this issue · comments
Thanks for interested in our work.
For the Rugby dataset, you have to prepare following data.
- Court model. It is the geometry information of the Rugby field. For example, the white lines shows https://en.wikipedia.org/wiki/Rugby_league_playing_field. These 2D lines are the geometry model that will be projected to image using simulated camera views.
- Training data: for each training image, there is a camera pose (location/orientation) with respect to the world coordinate. For example, the origin of the world coordinate is the most left-bottom corner of the court model.
- Statistics of the camera in the training data. For example, its mean location and standard deviation. These information are used to sample simulated camera views.
With these data, we can start the training process.
Awesome, thanks very much for this prompt and stepwise clear explanation for Rugby input requirement.
One more question, Can you please explain how these trained .mat files are generated as there isn't any code provided for training from scratch.
Looking forward for your response. Thank you once more.
Sure, but it is not clear to me which .mat file you refer to. Do you mean some .mat files in https://github.com/lood339/SCCvSD/tree/master/data? There is a readme file for the data https://github.com/lood339/SCCvSD/blob/master/data/read_me.py
There are two train steps in the pipeline.
The first one is to train a network to detect lines/circles from image. The input is an image. The output is an edge image. Details are in https://github.com/lood339/pytorch-two-GAN.
The second one is to train a siamese network. The input is an edge image. The output is a deep feature. The purpose is to reduce the dimension of edge image. Details are in https://github.com/lood339/SCCvSD/tree/master/python/deep