This is the official repository of the seminar paper Crowd counting with machine learning techniques, which is part of the IBB master's course at University of Ljubljana, Slovenia. You can read the report here.
In the paper we analyzed 5 CNN models for crowd counting and combined two to make our improvement. We provide the code of the improved model in a separate repository and we link the official implementations and papers of other models.
This repository contains the paper, links to the used models, and link to the datasets.
In our paper we analyze:
- CSRNet (paper, official implementation)
- Bayesian Crowd Counting (paper, official implementation)
- DM-Count (paper, official implementation)
- SFA-Net (paper, official implementation)
- SGANet (paper, official implementation)
Our improvement of the CSRNet:
- Bayesian CSRNet (implementation)
We provide the pretrained weights for the mentioned models on Google drive, so you can test and evaluate them yourself. Note that you might have to fix some data paths in the official implementations.
In this paper we evaluate the models on ShanghaiTech part A and part B datasets (download), as well as UCF-QNRF dataset (download).
We also provide an online interactive demo on Heroku. Please bear in mind that the demo uses CPU for evaluation, and due to the Heroku limitations can't process large images.
If you use our model or any of the models described in our paper, or the mentioned datasets, please cite them accordingly.