Crowd Counting Package

crowdcount is a library for crowd counting with Pytorch and supported by Fudan-VTS Research

Source

• code: https://github.com/FDU-VTS/crowd-count
• document: https://crowd-count.readthedocs.io/en/latest/

Install

• pip install crowdcount --user --upgrade

Introduction

Crowd counting task:

• estimate the number of crowd
• 

User guide:

• models

   from crowdcount.models import * 
   # crowd counting models includes csr_net, mcnn, resnet50, resnet101, unet, vgg
  

• transforms

   import crowdcount.transforms as cc_transforms
   # transforms
  

• data_loader

   from crowdcount.data.data_loader import *
   # includes ShanghaiTech, UCF_QNRF, UCF_CC_50, Fudan-ShanghaiTech temporarily
  

• data_preprocess

   from crowdcount.data.data_preprocess import *
   # gaussian preprocess for datasets
  

• utils

   from crowdcount.utils import *
   # includes loss functions, optimizers, tensorboard and save function
  

• engine

   from crowdcount.engine import train
   # start to train
   train(*args, **kwargs)
  

• More details in document

Demo

from crowdcount.engine import train
from crowdcount.models import Res101
from crowdcount.data.data_loader import *
from crowdcount.utils import *
import crowdcount.transforms as cc_transforms
import torchvision.transforms as transforms

# init model
model = Res101()
# init transforms
img_transform = transforms.Compose([transforms.ToTensor(),
                                    transforms.Normalize(mean=[0.452016860247, 0.447249650955, 0.431981861591],
                                                         std=[0.23242045939, 0.224925786257, 0.221840232611])
                                    ])
gt_transform = cc_transforms.LabelEnlarge()
both_transform = cc_transforms.ComplexCompose([cc_transforms.TransposeFlip()])
# init dataset
train_set = ShanghaiTechDataset(mode="train",
                                part="b",
                                img_transform=img_transform,
                                gt_transform=gt_transform,
                                both_transform=both_transform,
                                root="/home/vts/chensongjian/CrowdCount/crowdcount/data/datasets/shtu_dataset_sigma_15")
test_set = ShanghaiTechDataset(mode="test",
                               part='b',
                               img_transform=img_transform,
                               root="/home/vts/chensongjian/CrowdCount/crowdcount/data/datasets/shtu_dataset_sigma_15")
# init loss
train_loss = AVGLoss()
test_loss = EnlargeLoss(100)
# init save function
saver = Saver(path="../exp/2019-12-22-main_sigma15_6")
# init tensorboard
tb = TensorBoard(path="../runs/2019-12-22-main_sigma15_6")
# start to train
train(model, train_set, test_set, train_loss, test_loss, optim="Adam", saver=saver, cuda_num=[3], train_batch=2,
      test_batch=2, learning_rate=1e-5, epoch_num=500, enlarge_num=100, tensorboard=tb)

• you can find more demos in demo

Experiments

we will add the results soon

Thanks for the supports from

• C-3-Framework
• CSRNet
• Awesome-Crowd-Counting