Project: Deep-Feature-Flow-Segmentation (GitHub Link)

Deep-Feature-Flow-Segmentation-master
- init.sh
- LICENSE
- lib
  - utils
    - lr_scheduler.py
    - show_masks.py
    - show_boxes.py
    - image.py
    - mask_voc2coco.py
    - save_model.py
    - load_data.py
    - symbol.py
    - roidb.py
    - image_processing.py
    - combine_model.py
    - create_logger.py
    - mask_coco2voc.py
    - show_offset.py
    - __init__.py
    - tictoc.py
    - network_visualization.py
    - load_model.py
    - PrefetchingIter.py
  - Makefile
  - segmentation
    - __init__.py
    - segmentation.py
  - dataset
    - coco.py
    - imdb.py
    - cityscape.py
    - pascal_voc_eval.py
    - cityscape_video.py
    - pycocotools
      - coco.py
      - UPSTREAM_REV
      - maskApi.h
      - mask.py
      - setup_windows.py
      - cocoeval.py
      - maskApi.c
      - setup_linux.py
      - __init__.py
      - _mask.pyx
      - .gitignore
    - __init__.py
    - pascal_voc.py
    - ds_utils.py
  - nms
    - gpu_nms.pyx
    - gpu_nms.hpp
    - cpu_nms.pyx
    - nms_kernel.cu
    - setup_windows.py
    - setup_linux.py
    - __init__.py
    - setup_windows_cuda.py
    - nms.py
  - bbox
    - bbox_regression.py
    - setup_windows.py
    - setup_linux.py
    - __init__.py
    - .gitignore
    - bbox.pyx
    - bbox_transform.py
  - mask
    - mask_transform.py
    - __init__.py
  - __init__.py
  - logger
    - visdomlogger.py
    - logger.py
    - __init__.py
    - readme.md
  - rpn
    - generate_anchor.py
    - rpn.py
    - __init__.py
- DockerConfigs
  - README.md
  - MXNet
    - install
      - source.list
      - cpp.sh
      - python.sh
    - Dockerfile.python.gpu
- experiments
  - deeplab_dff
    - deeplab_dff_test.py
    - deeplab_dff_train.py
    - cfgs
      - deeplab_resnet_v1_101_cityscapes_segmentation_video_duc.yaml
      - deeplab_resnet_v1_101_cityscapes_segmentation_video.yaml
  - faster_rcnn
    - rcnn_test.py
    - rcnn_end2end_train_test.py
    - rcnn_train_test.py
    - cfgs
      - resnet_v1_101_voc0712_rcnn_dcn_end2end.yaml
      - resnet_v1_101_voc0712_rcnn_end2end.yaml
      - resnet_v1_101_coco_trainval_rcnn_dcn_end2end.yaml
      - resnet_v1_101_coco_trainval_rcnn_end2end.yaml
  - rfcn
    - rfcn_train_test.py
    - rfcn_alternate_train_test.py
    - rfcn_test.py
    - rfcn_end2end_train_test.py
    - cfgs
      - resnet_v1_101_voc0712_rfcn_end2end_ohem.yaml
      - deform_conv_demo.yaml
      - resnet_v1_101_coco_trainval_rfcn_dcn_end2end_ohem.yaml
      - resnet_v1_101_voc0712_rfcn_dcn_end2end_ohem.yaml
      - resnet_v1_101_coco_trainval_rfcn_end2end_ohem.yaml
      - deform_psroi_demo.yaml
      - rfcn_coco_demo.yaml
  - deeplab
    - deeplab_test.py
    - deeplab_train_test.py
    - cfgs
      - deeplab_resnet_v1_101_voc12_segmentation_dcn.yaml
      - deeplab_resnet_v1_101_voc12_segmentation_base.yaml
      - deeplab_resnet_v1_101_cityscapes_segmentation_capsule.yaml
      - deeplab_cityscapes_demo.yaml
      - deeplab_resnet_v1_101_cityscapes_segmentation_dcn.yaml
      - deeplab_resnet_v1_101_cityscapes_segmentation_base.yaml
- data
  - cityscapes
- README.md
- requirements.txt
- model
  - pretrained_model
    - resnet_v1_101-0000.params
- DeeplabCityscapes.tgn
- .gitignore
- deeplab
  - _init_paths.py
  - dff_train.py
  - function
    - test_deeplab.py
    - __init__.py
    - reeval.py
  - demo.py
  - config
    - config.py
    - dff_config.py
    - __init__.py
  - dff_test.py
  - train.py
  - __init__.py
  - core
    - module.py
    - tester.py
    - metric.py
    - callback.py
    - __init__.py
    - DataParallelExecutorGroup.py
    - loader.py
  - test.py
  - symbols
    - resnet_v1_101_deeplab.py
    - deeplabv3
      - aspp_temp.py
      - xception_temp.py
      - xception_65_deeplab_v3_plus.py
    - print_summary.py
    - resnet_v1_101_deeplab_video_dcn.py
    - resnet_v1_101_deeplab_dcn.py
    - resnet_v1_101_deeplab_video.py
    - __init__.py
    - resnet_v1_101_deeplab_dcn_duc.py

Deep Feature Flow for Video Semantic Segmentation

Based on Deeplab V2

1. Setup environment

If you use our dockerfile, you can run the code easily.
If you want to set up your own env, please follow these steps:
- We only support python2.7 now
- Install tk: sudo apt-get -y install python-tk
- Install OpenCV 3.4.1
- Install needed python packages with pip install -r requirements.txt
  - If you are in China Mainland, you can use these to speedup pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
Then sh init.sh to build the lib for faster-rcnn Because we use the code from Deformable ConvNets and the dataloader has some dependencies on faster-rcnn, so you need to build the lib first.
2. Prepare Data and Pretrained Model

Cityscapes Data

You need to download the cityscapes data from the official webpapge and unzip the data Put the data into data/cityscapes, you can use soft link to set the data path as the following: ln -s Dataset_path ./data/cityscapes

If you want to try DFF, you should download cityscapes video data and put it into data/cityscapes_video

Pretrained Model

Download pretrained resnet model flow net from Onedrive, and put the model into mode/pretrained_model/

./model/pretrained_model/resnet_v1_101-0000.params
./model/pretrained_model/flownet-0000.params

3. Train and Test

Training Deeplab V2

python ./experiments/deeplab/deeplab_train_test.py --cfg ./experiments/deeplab/cfgs/deeplab_resnet_v1_101_cityscapes_segmentation_base.yaml

Training Deeplab V2 Deformable

python ./experiments/deeplab/deeplab_train_test.py --cfg ./experiments/deeplab/cfgs/deeplab_resnet_v1_101_cityscapes_segmentation_dcn.yaml

Training DFF Deeplab V2

python ./experiments/deeplab_dff/deeplab_dff_train.py --cfg ./experiments/deeplab_dff/cfgs/deeplab_resnet_v1_101_cityscapes_segmentation_video.yaml

4. Performance

TBD

5. TODO List

[x] Add Scripts
[ ] Add experiment results
[ ] Add support for Deeplab V3+
[ ] Add BiSeNet
6. FAQ
Program hang if your system opencv is 2.x and your opencv-python is 3.x

7. Acknowledgement

Thanks for the official deep featuere flow implementation and deeplab implementation from MSRACVER