bowenc0221 / boundary-iou-api

Boundary IoU API (Beta version)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Boundary IoU API (Beta version)

Bowen Cheng, Ross Girshick, Piotr Dollár, Alexander C. Berg, Alexander Kirillov

[arXiv] [Project] [BibTeX]

This API is an experimental version of Boundary IoU for 5 datasets:

To install Boundary IoU API, run:

pip install git+https://github.com/bowenc0221/boundary-iou-api.git

or

git clone git@github.com:bowenc0221/boundary-iou-api.git
cd boundary_iou_api
pip install -e .

OpenCV is required to use Boundary IoU API.

Summary of usage

We provide two ways to use this api, you can either replace imports with our api or do offline evaluation.

Replacing imports

Our Boundary IoU API supports both evaluation with Mask IoU and Boundary IoU with the same interface as original ones. Thus, you only need to change the import, without worried about breaking your existing code.

  1. COCO instance segmentation
    replace

    from pycocotools.coco import COCO
    from pycocotools.cocoeval import COCOeval

    with

    from boundary_iou.coco_instance_api.coco import COCO
    from boundary_iou.coco_instance_api.cocoeval import COCOeval

    and set

    COCOeval(..., iouType="boundary")
  2. LVIS instance segmentation
    replace

    from lvis import LVISEval

    with

    from boundary_iou.lvis_instance_api.eval import LVISEval

    and set

    LVISEval(..., iou_type="boundary")
  3. Cityscapes instance segmentation
    replace

    import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as cityscapes_eval

    with

    import boundary_iou.cityscapes_instance_api.evalInstanceLevelSemanticLabeling as cityscapes_eval

    and set

    cityscapes_eval.args.iou_type = "boundary"
  4. COCO panoptic segmentation
    replace

    from panopticapi.evaluation import pq_compute

    with

    from boundary_iou.coco_panoptic_api.evaluation import pq_compute

    and set

    pq_compute(..., iou_type="boundary")
  5. Cityscapes panoptic segmentation
    replace

    from cityscapesscripts.evaluation.evalPanopticSemanticLabeling as evaluatePanoptic

    with

    from boundary_iou.cityscapes_panoptic_api.evalPanopticSemanticLabeling import evaluatePanoptic

    and set

    evaluatePanoptic(..., iou_type="boundary")

Offline evaluation

We also provide evaluation code that can evaluates your prediction files for each dataset.

  1. COCO instance segmentation

    python ./tools/coco_instance_evaluation.py \
        --gt-json-file COCO_GT_JSON \
        --dt-json-file COCO_DT_JSON \
        --iou-type boundary
  2. LVIS instance segmentation

    python ./tools/lvis_instance_evaluation.py \
        --gt-json-file LVIS_GT_JSON \
        --dt-json-file LVIS_DT_JSON \
        --iou-type boundary
  3. Cityscapes instance segmentation

    python ./tools/cityscapes_instance_evaluation.py \
        --gt_dir GT_DIR \
        --result_dir RESULT_DIR \
        --iou-type boundary
  4. COCO panoptic segmentation

    python ./tools/coco_panoptic_evaluation.py \
        --gt_json_file PANOPTIC_GT_JSON \
        --gt_folder PANOPTIC_GT_DIR \
        --pred_json_file PANOPTIC_PRED_JSON \
        --pred_folder PANOPTIC_PRED_DIR \
        --iou-type boundary
  5. Cityscapes panoptic segmentation

    python ./tools/cityscapes_panoptic_evaluation.py \
        --gt_json_file PANOPTIC_GT_JSON \
        --gt_folder PANOPTIC_GT_DIR \
        --pred_json_file PANOPTIC_PRED_JSON \
        --pred_folder PANOPTIC_PRED_DIR \
        --iou-type boundary

Citing Boundary IoU

If you find Boundary IoU helpful in your research or wish to refer to the referenced results, please use the following BibTeX entry.

@inproceedings{cheng2021boundary,
  title={Boundary {IoU}: Improving Object-Centric Image Segmentation Evaluation},
  author={Bowen Cheng and Ross Girshick and Piotr Doll{\'a}r and Alexander C. Berg and Alexander Kirillov},
  booktitle={CVPR},
  year={2021}
}

Contact

If you have any questions regarding this API, please contact us at bcheng9 AT illinois.edu

About

Boundary IoU API (Beta version)

License:Other


Languages

Language:Python 100.0%