Paper

Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, Chen Change Loy, "Self-Supervised Scene De-occlusion", accepted to CVPR 2020 as an Oral Paper. [Project page].

For further information, please contact Xiaohang Zhan.

Demo Video

• Watch the full demo video in YouTube or bilibili. The demo video contains vivid explanations of the idea, and interesting applications.

• Below is an application of scene de-occlusion: image manipulation. Code: deocclusion-demo

Requirements

• python: 3.7

• pytorch>=0.4.1

• install pycocotools:

  pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"

• others:

  pip install -r requirements.txt

Run Demos

1. Download released models here and put the folder released under deocclusion.

2. Run demos/demo_cocoa.ipynb or demos/demo_kins.ipynb. There are some test examples for demos/demo_cocoa.ipynb in the repo, so you don't have to download the COCOA dataset if you just want to try a few samples.

3. If you want to use predicted modal masks by existing instance segmentation models, you need to adjust some parameters in the demo, please refer to the answers in this issue.

Data Preparation

COCOA dataset proposed in Semantic Amodal Segmentation.

1. Download COCO2014 train and val images from here and unzip.

2. Download COCOA annotations from here and untar.

3. Ensure the COCOA folder looks like:

   COCOA/
     |-- train2014/
     |-- val2014/
     |-- annotations/
       |-- COCO_amodal_train2014.json
       |-- COCO_amodal_val2014.json
       |-- COCO_amodal_test2014.json
       |-- ...
   

4. Create symbolic link:

   cd deocclusion
   mkdir data
   cd data
   ln -s /path/to/COCOA
   

KINS dataset proposed in Amodal Instance Segmentation with KINS Dataset.

1. Download left color images of object data in KITTI dataset from here and unzip.

2. Download KINS annotations from here corresponding to this commit.

3. Ensure the KINS folder looks like:

   KINS/
     |-- training/image_2/
     |-- testing/image_2/
     |-- instances_train.json
     |-- instances_val.json
   

4. Create symbolic link:

   cd deocclusion/data
   ln -s /path/to/KINS
   

LVIS dataset

1. Download training and validation sets from here

Using your own dataset

If using your own dataset to train or test, you need to make sure that it contains accurate modal annotations (masks are required and categories are optional). Inaccurate modal mask annotations, e.g., COCO original annotaions that may have large margin between masks of occluding objects, will result in unsatisfactory results.

Train

train PCNet-M

1. Train (taking COCOA for example).

   sh experiments/COCOA/pcnet_m/train.sh # you may have to set --nproc_per_node=#YOUR_GPUS
   

2. Monitoring status and visual results using tensorboard.

   sh tensorboard.sh $PORT
   

train PCNet-C

1. Download the pre-trained image inpainting model using partial convolution here to pretrains/partialconv.pth

2. Convert the model to accept 4 channel inputs.

   python tools/convert_pcnetc_pretrain.py

3. Train (taking COCOA for example).

   sh experiments/COCOA/pcnet_c/train.sh # you may have to set --nproc_per_node=#YOUR_GPUS
   

4. Monitoring status and visual results using tensorboard.

Evaluate

• Execute:

  sh tools/test_cocoa.sh

Bibtex

@inproceedings{zhan2020self,
 author = {Zhan, Xiaohang and Pan, Xingang and Dai, Bo and Liu, Ziwei and Lin, Dahua and Loy, Chen Change},
 title = {Self-Supervised Scene De-occlusion},
 booktitle = {Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR)},
 month = {June},
 year = {2020}
}

Acknowledgement

1. We used the code and models of GCA-Matting in our demo.

2. We modified some code from pytorch-inpainting-with-partial-conv to train the PCNet-C.