IUML: INCEPTION U-NET BASED MULTI-TASK LEARNING FOR DENSITY LEVEL CLASSIFICATION AND CROWD DENSITY ESTIMATION

By Van-Su Huynh, Vu-Hoang Tran, and Chin-Chung Huang

This work is accepted by SMC 2019 conference

Introduction

This project is an implementation of IUML network for crowd counting. IUML network could handle various types of scale problem caused by: depth variation, height variation, the variation caused by density levels, and image resolution difference

Dependencies and Installation

We have tested the implementation on Window with GPU Nvidia 1080TI, CUDA8 and CuDNN v5 . The other version should be working. Caffe installation is pre-required.

Dataset preparation

The ShanghaiTech dataset (1) could be dowloaded here. The UCF_CC_50 dataset (2) could be dowloaded here.

After getting the dataset, using the codes in data_preparation to create the training patch. Each original image, we randomly generate 30 patches. We applied a geometry-adaptive kernel (1) which results in a smaller kernel size for a smaller object and a larger kernel size for a larger object.

Training and Test

The hyper-parameters are denoted in file_solver.prototxt The training model was defined in file_train.prototxt The testing model was defined in deploy.prototxt

References

(1) Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma, “Single-image crowd counting via multi-column convolutional neural network,” CVPR, 2016.

(2) Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak Shah, “Multi-source multi-scale counting in extremely dense crowd images,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2547-2554, 2013.