Efficient-Uncertainty-Video-Segmentation

This is the official codes for the paper: Efficient Uncertainty Estimation for Semantic Segmentation in Videos.

Requirements

• Python 2.7
• Pytorch 0.2.0
• tqdm
• matplotlib
• Visdom 0.1.7
• pypng
• protobuf
• Opencv

CamVid dataset

Normall CamVid dataset only contain frames that fps=1.

However, our Method leverage consecutive frames to speed up uncertainty estimation.

Therefore we need a CamVid dataset contain all consecutive frames(fps 30) instead of labeled frames(fps 1).

We extract fps 30 frames from original videos and build new version here

Download and unzip the dataset Edit config.json

{
  "camvid":
  {
    "data_path": "/YOUR/PATH/camvid/"
  }
}

Optical flow installation

We use the FlowNet2 as our optical flow model.

The FlowNet2 code is intergret in our repo.

You only need do the installation in this repo FlowNet2.

Download the pretrained weight in Dir : pytorch_flownet2/FlowNet2_src/pretrained

Trained model

Our trained tiramisu model can be download here.

Download and unzip it at checkpoint dir. Then run exp_test_MC.py and python exp_test_RTA.py.

Then it can evaluate our release model.

Train script

• Tiramisu

  python exp_train.py
  

Evaluate Script

• Tiramisu MC dropout (sample 5 times) Important hyper-parameter

  mode = 'test'
  ckpt_epoch = 900
  video_unct = False
  sample_num = 5
  

  Command

  python exp_test_MC.py
  

• Tiramisu TA-MC Important hyper-parameter

  mode = 'test'
  ckpt_epoch = 900
  video_unct = False
  error_thres = 300 
  alpha_normal = 0.2
  alpha_error = 0.7
  

  Command

  python exp_test_RTA.py
  

• Tiramisu RTA-MC Important hyper-parameter

  mode = 'test'
  ckpt_epoch = 900
  video_unct = False
  error_thres = 40
  alpha_normal = 0.2
  alpha_error = 0.7
  

  Command

  python exp_test_RTA.py
  

Results

• Tiramisu MC dropout N=5(we use N=5 result because the same inference time as RTA-MC.)

  • Performance

    	Accuracy
    Global Accuracy	89.3
    Mean Accuracy	75.3
    Mean IOU	62.6

  • PR-Curve

    

  • Ranking IOU of Variational Ratio

    Percentage	Ranking IOU
    10%	43.4
    30%	58.0
    50%	73.3
    70%	85.2

• Tiramisu TA-MC

  • Performance

    	Accuracy(%)
    Global Accuracy	89.6
    Mean Accuracy	73.5
    Mean IOU	62.2

  • PR-Curve

    

  • Ranking IOU of Variational Ratio

    Percentage	Ranking IOU
    10%	34.8
    30%	60.9
    50%	76.7
    70%	87.0

• Tiramisu RTA-MC

  • Performance

    	Accuracy(%)
    Global Accuracy	89.6
    Mean Accuracy	74.2
    Mean IOU	62.6

  • PR-Curve

    

  • Ranking IOU of Variational Ratio

    Percentage	Ranking IOU
    10%	43.4
    30%	65.2
    50%	77.6
    70%	86.4