UNet segmentation metrics

WORK IN PROGRESS

Simple Python 3 tools to assess the performance of UNet segmentation networks (or any other segmentation method) by comparing the prediction to a ground truth image.

Use it to calculate:

• Jaccard metric for object detection
• Intersection over Union (IoU) for object segmentation accuracy
• Localization (positional) error for estimating MOTP during tracking
• Pixel identity

TODO:

• Add strict matching with IoU threshold
• Add confusion matrix for multi-label/classification type tasks

Single image usage

import umetrics
from skimage.io import imread

y_true = imread('true.tif')
y_pred = imread('pred.tif')


# can now make the calculation strict, by only considering objects that have
# an IoU above a theshold as being true positives
result = umetrics.calculate(y_true, y_pred, strict=True, iou_threshold=0.5)

print(result.results)

returns:

============================
 Segmentation Metrics (n=1)
============================
Strict: True (IoU > 0.5)
n_true_labels: 354
n_pred_labels: 362
n_true_positives: 345
n_false_positives: 10
n_false_negatives: 0
IoU: 0.999
Jaccard: 0.972
pixel_identity: 0.998
localization_error: 0.010

Batch processing

import umetrics

# provide a list of file pairs ('true', 'prediction')
files = [('true0.tif', 'pred0.tif'),
         ('true1.tif', 'pred1.tif'),
         ('true2.tif', 'pred2.tif')]

batch_result = umetrics.batch(files)

Returns aggregate statistics over the batch. Jaccard index is calculated over all found objects, while other metrics are the average IoU etc.

Installation

1. First clone the repo:

$ git clone https://github.com/quantumjot/unet_segmentation_metrics.git

2. (Optional, but advised) Create a conda environment:

$ conda create -n umetrics python=3.7
$ conda activate umetrics

3. Install the package

$ cd unet_segmentation_metrics
$ pip install .