What’s Outside the Intersection? Fine-grained Error Analysis for Semantic Segmentation Beyond IoU

This is the code used to produce the results in the paper.

Getting started

• install mmsegmentation>=1.0.0 with all its requirements (e.g. numpy, torch)
• install tqdm
• [optional] install this repository locally with pip install -e .

Usage

To run the error analysis, only the path to ground-truth and predictions (stored as images) as well as some metainfo is necessary, e.g.:

python eval.py \
    --pred_dir=path/to/predictions \
    --gt_dir=path/to/gt \
    --dataset_name=ADE20KDataset \
    --output_dir=path/to/result \
    --ignore_index=255 \
    --boundary_width=0.01 \
    --reduce_zero_label # this maps label 0 to 255, 1 to 0 etc.

Alternatively, one can create the evaluator (as well as the loader) manually, as done in the main method of eval.py.

Result

An example for the results obtained with the error analysis can be found in example_result_mask2former_swin-l_ade20k.