Multi-Object-Occlusion

Robust Instance Segmentation through Reasoning about Multi-Object Occlusion [CVPR 2021]

Abstract

Analyzing complex scenes with DNN is a challenging task, particularly when images contain multiple objects that partially occlude each other. Existing approaches to image analysis mostly process objects independently and do not take into account the relative occlusion of nearby objects. We propose a deep network for multi-object instance segmentation that is robust to occlusion and can be trained from bounding box supervision only.

We also introduce an Occlusion Challenge dataset generated from real-world segmented objects with accurate annotations and propose a taxonomy of occlusion scenarios that pose a particular challenge for computer vision.


NOTICE

dataset links and model will be released in a few days. Update: 18 June

Requirments

The code uses Python 3.6 and it is tested on PyTorch GPU version 1.2, with CUDA-10.0 and cuDNN-7.5.

Installation

  1. Clone the repository with:

    git clone https://github.com/XD7479/Multi-Object-Occlusion.git
    cd Multi-Object-Occlusion

  2. Install requirments:

    pip install -r requirements.txt

Datasets

  1. Download the KINS dataset here and the Occlusion Challenge dataset here.

  2. Enter the project folder and make links for the datasets:

    ln -s <YOUR_KINS_DATASET_PATH> kins
    ln -s <YOUR_OCC_CHALLENGE_DATASET_PATH> occ_challenge

  3. Download the pre-trained model here.

  4. Make links for the pre-trained model:

    ln -s <YOUR_MODEL_PATH> models

  5. Check the configuration file configs.py for the dataset and backbone you're using:

    dataset_eval = 'occ_challenge' # kins, occ_challenge
    nn_type = 'resnext' # vgg, resnext

  6. Run the evaluation code with:

    python3 eval_meanIoU.py


Segmentation Demo

visual_seg_synth

Citation

@misc{yuan2021robust,
      title={Robust Instance Segmentation through Reasoning about Multi-Object Occlusion}, 
      author={Xiaoding Yuan and Adam Kortylewski and Yihong Sun and Alan Yuille},
      booktitle = {Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
      month = jun,
      year = {2021},
      month_numeric = {6}
}

Contact

If you have any questions you can contact Xiaoding Yuan by [email protected].

GitHub

https://github.com/XD7479/Multi-Object-Occlusion