Khoa-NT / rec-attend-public

Code that implements paper "End-to-End Instance Segmentation with Recurrent Attention"

rec-attend-public

Code that implements paper "End-to-End Instance Segmentation with Recurrent Attention".

Dependencies

• Python 2.7
• TensorFlow 0.12 (not compatible with TensorFlow 1.0)
• OpenCV
• NumPy
• SciPy
• PyYaml
• hdf5 and H5Py
• tqdm
• Pillow (required by cityscapes evaluation)

Installation

Compile Hungarian matching module

./hungarian_build.sh

CREMI Experiments

Note: Unlock h5_lock:
export HDF5_USE_FILE_LOCKING='FALSE'

Crop data:

Run cremi_Prepare_Eval.ipynb

Setup data:

Configure the size opt in setup_cvppp.py.
Note: setup_cvppp.py will automatically resize to the size opt in setup_cvppp.py.

Run python setup_cvppp.py

###Run experiments:

• Configure the setting:

  • Class TrainArgsParser in cmd_args_parser.py
  • Number of object kCvpppNumObj in cmd_args_parser.py
  • return in get_default_timespan in cvppp.py. Have to return = kCvpppNumObj+1
  • steps_per_valid in cmd_args_parser.py
  • steps_per_trainval in cmd_args_parser.py
  • steps_per_plot in cmd_args_parser.py
  • num_batch_valid in cmd_args_parser.py
  • MAX_NUM_ITERATION in hungarian.cc to prevent core_dumped
  • add --fixed_order to turn off Hungarian

• Choose GPU_id in:

  • box_model_train.py
  • box_model_read.py
  • full_model_train.py
  • full_model_eval.py

• Comment those code in box_model_train.py

    # if 'attn' in self.loggers:
    #   pu.plot_double_attention(
    #       self.loggers['attn'].get_fname(),
    #       x,
    #       results['ctrl_rnn_glimpse_map'],
    #       max_items_per_row=max_items)
  

Run

./run_cremi.sh

CVPPP Experiments

First modify setup_cvppp.sh with your dataset folder paths.

./setup_cvppp.sh

Run experiments:

./run_cvppp.sh

KITTI Experiments

First modify setup_kitti.sh with your dataset folder paths.

./setup_kitti.sh

Run experiments:

./run_cvppp.sh

Cityscapes Experiments

First modify setup_cityscapes.sh with your dataset folder paths.

./setup_cityscapes.sh

Run experiments:

./run_cityscapes.sh

Citation

If you use our code, please consider cite the following: End-to-End Instance Segmentation with Recurrent Attention. Mengye Ren, Richard S. Zemel. CVPR 2017.

@inproceedings{ren17recattend,
  author    = {Mengye Ren and Richard S. Zemel},
  title     = {End-to-End Instance Segmentation with Recurrent Attention},
  booktitle = {CVPR},
  year      = {2017}
}