An improved implementation of CSRNet using TensorFlow, focusing on analysis of crowd patterns in densely populated images via a completely convolutional model.

CSRNet is an innovative Deep Learning model proposed by Yuhong Li et al. with an objective to understand the crowd pattern in densely populated images. The model is a fully convolutional with a FrontEnd and a BackEnd.

In this project, we have successfully implemented all the four architectures as mentioned in the CSRNet paper. The easy-to-use Keras API of TensorFlow served as our primary tool for this implementation.

The TensorFlow's Keras API provides a multitude of high-level operations ensuring simplicity in coding without compromising the performance. The Input pipeline employs an efficient Data API pipeline that parses tfRecord files.

For this, we generate a tfRecord file with all the required input data.

The first 13 layers of the VGGNet pre-trained ImageNet were used as our FrontEnd Network. The BackEnd, on the other hand, is built with different combination of Convolution and Dilated Convolutions primary named as A, B, C, and D.

The primary Loss function for this implementation was the simple L2 loss, as we use smoothed label as the target.

We used the Adam Optimizer to train our model. The first 13 layers of the pre-trained VGGNet were frozen, allowing only the Backend network to be trained effectively.

Training makes use of the ShangaiTech Dataset and spans over 1,70,000 iterations which roughly takes around 36 hours on a system equipped with a NVIDIA 1080Ti graphics card and an Intel i5 processor paired with a 16GB RAM.

Refer to the GIFs below for a better understanding of our model's performance. The model was successful in capturing the crowd patterns in majority of the images.

Test Images

Ground Truth Labels

Our Predictions

Continuous optimization in training can lead to better results.