Synthesize then Compare: Detecting Failures and Anomalies for Semantic Segmentation

Project page | Paper | Video

Synthesize then Compare: Detecting Failures and Anomalies for Semantic Segmentation
Yingda Xia*, Yi Zhang*, Fengze Liu, Wei Shen, and Alan Yuille.
In ECCV 2020 (Oral).

Installation

Clone this repo.

git clone https://github.com/YingdaXia/SynthCP.git
cd SynthCP

This code has been tested with PyTorch 1.2.0 and python 3.7. Please install dependencies by

pip install -r requirements.txt

Dataset Preparation

Download Cityscapes dataset and place it under datasets/.

Download StreetHards dataset following this repo. Please download both train and test data, and arrange them as anomaly/data/train and anomaly/data/test respectively.

Running the code on the Cityscpaes

First, download our checkpoint models from here and name the folder as `checkpoints/'. You should have a structure like this,

checkpoints/
  |----fcn8s/
  |----deeplab/
  |----iounet/
  |----cityscapes_label_only_c19/
  |----caos/

Obtaining paper results

Take FCN-8s as example, you will need to test the FCN model on the testset of Cityscapes. See scripts/eval_segmentation_models.sh.

bash scripts/eval_segmentation_models.sh

Then run the scripts we provided to obtain numbers.

• SynthCP: scripts/reproduce_synthcp_{fcn,deeplab}.sh
• Direct Prediction: scripts/reproduce_synthcp_{fcn,deeplab}.sh
• MSP: scripts/reproduce_synthcp_{fcn,deeplab}.sh
• MCDropout: scripts/reproduce_mcdropout_{fcn,deeplab}.sh
• TCP: scripts/reproduce_tcp.sh

Training

To train the synthesize module (SPADE), use the following script:

cd spade-cityscapes
bash run.sh

We provided our pre-trained SPADE (on the label-image pairs) in checkpoints/cityscapes_label_only_c19/.

To train the comparison module, you need to first train the segmentation models using cross-validation on Cityscapes training set,

bash scripts/train_segmentation_models.sh

Then evaluate the model on the validation set of each fold (See the commented lines in scripts/eval_segmentation_models.sh).

bash scripts/eval_segmentation_models.sh

Also run the synthesize module on the training set (See the commented lines in scripts/reproduce_synthcp_{fcn,deeplab}.sh). Train the comparison module,

bash scripts/train_iounet.sh 0 $EXP_PATH $IOUNET_NAME $REC_PATH

Similar process for Direct Prediction, image-level MCDropout and TCP.

Running the code on StreetHazards dataset

First, train and test segmentation model to obtain segmentation predictions (saved in anomaly/data/test_result by default).

cd anomaly
python train.py
python test.py

Then, train the synthesize module (SPADE).

cd ../spade-caos
bash run.sh

And use it to obtain reconstructions of the segmentation predictions (saved in anomaly/data/test_recon by default).

bash eval_spade.sh

We also provided our trained model in checkpoints/caos. If you want to use it, please copy it to spade-caos/checkpoints/caos.

Finally, segment anomaly objects by computing a feature-space distance between the images and the reconstructions.

cd ../anomaly
python eval_ood_rec.py

Citation

If you use this code for your research, please cite our papers.

@inproceedings{xia2020synthesize,
  title={Synthesize then Compare: Detecting Failures and Anomalies for Semantic Segmentation},
  author={Xia, Yingda and Zhang, Yi and Liu, Fengze and Shen, Wei and Yuille, Alan},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2020}
}

Acknowledgments

The code for training and testing the synthesize module is extended from SPADE (Copyright (C) 2019 NVIDIA Corporation). The code for anomaly segmentation is extended from anomaly-seg, which is also built on semantic-segmentation-pytorch.