CSRNet-caffe

This is a caffe implementation of CSRNet for crowd counting.

Dependencies

matlab (2015b)
OpenCV (2.4.13)

How to use?

caffe build

Because add lmdb2txt tool to caffe

cd caffe-CSRNet
mkdir build
cd build
cmake ..
make -j8

prepare data

download Shanghai dataset

https://drive.google.com/file/d/16dhJn7k4FWVwByRsQAEpl9lwjuV03jVI/view

generate data list

1. cd readPicList
2. python3 readPicList.py -i ../ShanghaiTech_Crowd_Counting_Dataset/part_A_final/train_data/images -o train_list.txt 
3. python3 addSuffix.py --inputfile train_list.txt --outputfile_csv label.txt --outputfile_jpg img_train.txt

generate density map

copy readPicList/train_list.txt to generate_density folder and modify create_gt_train_set.m corresponding path, then run create_gt_train_set.m, the generated density map will be saved in ground_truth_csv folder.

convert ground_truth(.csv) to lmdb

note:modify the corresponding include and link path in CMakeList file eg./home/ozh/share/caffe/build/lib-->path/to/caffe/build/lib

cd csv2lmdb
mkdir build
cd build
make
cd ../bin
./csv2lmdb --lmdb_path /home/ozh/share/test/lmdb/ground_truth --csv_list ../../readPicList/label.txt --ground_truth_path ../../ShanghaiTech_Crowd_Counting_Dataset/part_A_final/train_data/ground_truth_csv/

convert data(.jpg) to lmdb

sh create_lmdb.sh (modify the corresponding path)

download pretrain model vgg16

http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel

train

sh train_CSRNet.sh (modify the corresponding path)

test

run crowd_test.m (modify the test_list.txt in deploy.prototxt)