lartpang/PySODMetrics

What it solves

PySODMetrics provides a standardized, efficient way to evaluate Salient Object Detection (SOD) and Camouflaged Object Detection (COD) models. It replaces the need for complex Matlab-based evaluation toolboxes with a lightweight Python implementation, ensuring consistent and verifiable performance measurement across different research papers.

How it works

Built on numpy and scipy, the library implements a wide array of segmentation metrics. It supports both sample-based and whole-based calculations, allowing researchers to evaluate models using various thresholding strategies (such as max, avg, adp, and bin). The implementation is verified against established Matlab benchmarks to ensure mathematical accuracy.

Who it’s for

Computer vision researchers and developers working on salient object detection, camouflaged object detection, or general image segmentation tasks who need reliable metrics to benchmark their models.

Highlights

• Extensive Metric Support: Includes MAE, S-measure, weighted F-measure, E-measure, Multi-Scale IoU, and standard metrics like Precision, Recall, and Dice.
• Lightweight Design: Fast execution with a simple code structure that is easy to extend.
• Verified Accuracy: Results are consistent with the widely used CODToolbox Matlab code.
• Flexible Computation: Supports various sample-based and whole-based variants, including size-invariance (si-) variants.