Fork from https://github.com/nqanh/pose_relocalization Fixes to old version code
- python=2.7
- tensorflow-gpu=1.15.0
- Keras 2.0.3
- more details in environment.yaml
- note: Do not use 30series GPU(Some unknown errors)
- create dataset in dataset/ as pickle file
- the data in the pickle is a tuple,
- tuple[0] are the images, the shape is [number_of_images,224,224,3]
- tuple[1] are the posesm the shape is [number_of_images,7]
cd main_keras
python train.py --gpu GPU_ID
- For example:
python train.py --gpu 0
- --> will train on GPU 0
- The parameters of the original repository are for the original version. Here write the required paths directly to the file.
cd main_keras
python predict.py
- Change in_scene_id, in_net_id, in_model_file_name in this file for different scene
Instead of calling the read pickle method in dota_io_cnn.py, the pickle read is used directly in the train. (Unknown error)
_tkinter.TclError: no display name and no $DISPLAY environment variable import matplotlib matplotlib.use('Agg') before import matplotlib.pyplot as plt
If you find this code useful in your research, please consider citing:
@article{nguyen6dof_lstmpose,
author = {Anh Nguyen and
Thanh{-}Toan Do and
Darwin G. Caldwell and
Nikos G. Tsagarakis},
title = {Real-Time Pose Estimation for Event Cameras with Stacked Spatial {LSTM}
Networks},
journal = {IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
year = {2019}
}