Home

Awesome

AudioVisualCrowdCounting

This is the code for AudioVisual crowd counting. To use the code you need to install PyTorch-1.0 and Python 3.7.

Dataset

We propose a new dataset for crowd counting, which is composed of around 2000 annotaed images token in different locations in China and each image corresponds to a 1 second audio clip and a density map. The images are in different illuminations. More details can be found in our paper Ambient Sound Helps: Audiovisual Crowd Counting in Extreme Conditions and you can download the dataset here. We also provide the original dot annotations here, please feel free to use it.

Train

  1. Download the dataset including images, audios and density maps. Unzip the files and put them into the same folder, for example, ./audio_visual_data and then switch DATA_PATH in datasets/AC/setting.py to audio_visual_data.
  2. Download the pretrained VGGish and put it into ./models/SCC_Model/ folder.
  3. To train a model using raw images, setting IS_NOISE to False and BLACK_AREA_RATIO to 0.
  4. To train a model using low-quality images (low illumination and noisy)， setting IS_NOISE to True, BLACK_AREA_RATIO to 0 and BRIGHTNESS to [0,1]. The parameter IS_RANDOM indicates whether BRIGHTNESS is a fixed value or a random number during traning. Details can be found in our paper.
  5. You can also change the settings in config.py, such as the name of the model.

Test

After training, you can run my_test.py to test the trained model. Note that in my_tester.py we also save the predicted density map, you should switch the path self.save_path to your own setting.

Acknowledgement

The repository is derived from C-3-Framework.

Citation

@article{hu2020,
  title={Ambient Sound Helps: Audiovisual Crowd Counting in Extreme Conditions},
  author={Di Hu and Lichao Mou and Qingzhong Wang and Junyu Gao and Yuansheng Hua and Dejing Dou and Xiao Xiang Zhu},
  journal={arXiv preprint},
  year={2020}
}