Crowd Counting - Group 14
This repo contains several .ipynb files in which we ran our experiments (Definining models, Preparing dataset, training, testing) You can run them locally or import them to Google Colab
A fork of an implementation of Multi Column Neural Network for Crowd counting. Contains definition of the model, dataloader and other files for test and training, plus a small modification on the model in which we try to improve its acccuracy. Later experiments with the MCNN's will be based on this architecture, but in a single .ipynb file instead.
One of our attempts to implement a new technique (Point-to-point estimation) into the original MCNN architecture. Due to implementation difficulties, lack of time and our priorization of other more promising approaches, we stopped working on it
This file includes the 2 final methods we proposed in our final presentation video. The first one is Multi-Column Convolutional Neural Network (MCNN) enhanced with a Distribution Matching (DM) technique and address the overfitting problem by data augmentation, drop neuron units, and decaying learning rate. The second method modifies the structure of CSRNet by adding the dilated kernel. It has the best performance in the ShanghaiTech Crowd dataset. You can reproduce the results by running the blocks one by one from top to bottom.