Encoder-Decoder Based Convolutional Neural Networks with Multi-Scale-Aware Modules for Crowd Counting（ICPR 2020）

Official Implementation of "Encoder-Decoder Based Convolutional Neural Networks with Multi-Scale-Aware Modules for Crowd Counting" LINK

Many thanks to BL, SFANet and CAN for their useful publications and repositories.

For complete UCF-QNRF and Shanghaitech training code, please refer to BL and SFANet respectively.

Please see models for our M-SFANet and M-SegNet implementations.

Density maps Visualization

Datasets (NEW)

To reproduce the results reported in the paper, you may use these preprocessed datasets. This is not completed yet, and might be updated in the future.

Shanghaitech B dataset that is preprocessed using the Gaussian kernel Link

Bayesian preprocessed (following BL) Shanghaitech datasets (A&B) Link

The Beijing-BRT dataset Link (Originally from BRT)

Pretrained Weights

Shanghaitech A&B Link

To test the visualization code you should use the pretrained M_SegNet* on UCF_QNRF Link (The pretrained weights of M_SFANet* are also included.)

Getting started

An example code of how to use the pretrained M-SFANet* on UCF-QNRF to count the number people in an image. The test image is ./images/img_0071.jpg (from UCF-QNRF test set).

import cv2
from PIL import Image
import numpy as np

import torch
from torchvision import transforms

from datasets.crowd import Crowd
from models import M_SFANet_UCF_QNRF

# Simple preprocessing.
trans = transforms.Compose([transforms.ToTensor(), 
                            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
                           ])

# An example image with the label = 1236.
img = Image.open("./images/img_0071.jpg").convert('RGB')
height, width = img.size[1], img.size[0]
height = round(height / 16) * 16
width = round(width / 16) * 16
img = cv2.resize(np.array(img), (width,height), cv2.INTER_CUBIC)
img = trans(Image.fromarray(img))[None, :]

model = M_SFANet_UCF_QNRF.Model()
# Weights are stored in the Google drive link.
# The model are originally trained on a GPU but, we can also test it on a CPU.
# For ShanghaitechWeights, use torch.load("./ShanghaitechWeights/...")["model"] with M_SFANet.Model() or M_SegNet.Model()
model.load_state_dict(torch.load("./Paper's_weights_UCF_QNRF/best_M-SFANet*_UCF_QNRF.pth", 
                                 map_location = torch.device('cpu')))

# Evaluation mode
model.eval()
density_map = model(img)
# Est. count = 1168.37 (67.63 deviates from the ground truth)
print(torch.sum(density_map).item())

Citation

If you find the code useful for your research, please cite our paper:

@inproceedings{thanasutives2021encoder,
  title={Encoder-decoder based convolutional neural networks with multi-scale-aware modules for crowd counting},
  author={Thanasutives, Pongpisit and Fukui, Ken-ichi and Numao, Masayuki and Kijsirikul, Boonserm},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)},
  pages={2382--2389},
  year={2021},
  organization={IEEE}
}

Erratum: In Fig. 1 of the paper, "ASSP" should be "ASPP".
You may watch this 6-minute presentation video as a short introduction.