People Counting Algorithm - CSRNet-pytorch

Good morning, I recently ordered a people counting algorithm. He referred to the implementation made in CSRNet-pytorch. Unfortunately he did not fully meet my initial requirement. What I need now is to train him from new images so that his accuracy is improved when the crowd is not too dense. I have 65 images with labels and would like some help generating a new training by adding to the dataset they provide these 65 images. As I understand it is necessary to generate before training a map (.mat file), a Json file containing information about what will be trained. I can forward the source code I have to adapt it if necessary and the 65 images to add to the set of images provided by CSRNet. thanks,