ASD-crowd-counting

This is an unofficial implementation of the ICASSP 2019 paper Adaptive Scenario Discovery for Crowd Counting by PyTorch. Different with the paper, I added some data augmentation methods that turn out to be effective. The data augmentation methods reference from this paper.

Prerequisites

Python: 3.5

PyTorch: 1.0.1

Code structure

density_map.py To generate the density map.

data.py Data preprocess and augmentation.

model.py The structure of the network.

logger.py Utility for logging on tensorboard.

train.py To train the model.

eval.py To evaluate the model.

Train

I have trained the model on ShanghaiTech part B. This is the training and testing logs on TensorBoard.

Average loss on training set

MAE on test set

MSE on test set

Results

MAE: 7.28 MSE: 11.85 on ShanghaiTech part B.

Download checkpoint: Google Drive