FCN-rLSTM model trained on webcamt dataset

Capstone project : Density map estimation & Vehicle Counting

---

Datasets

• WebCamT
  • https://github.com/Lotuslisa/WebCamT (Github)
  • https://www.citycam-cmu.com/ (Dataset url)

Model

• FCN-rLSTM
  • Shanghang Zhang, Guanhang Wu, Joao P Costeira, and Jose MF Moura. Fcn-rlstm: Deep spatio-temporal neural networks for vehicle counting in city cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3667–3676, 2017.
  • Implementation notes - dpernes/FCN-rLSTM

Details

Memory overload occurs when about 59000 pieces of data are loaded at once

• Save the preprocessed data as a .pickle file at once or divided into two depending on the capacity of each folder.
• For each epoch, all pickles are loaded and trained, and after training on the last pickle file, the training results are saved to a log file in the path set for checking the results.
• Training and test data can be modified as needed.

Contact

• hongil7626@gmail.com