A benchmark for environmental microorganism object detection on EMDS-7.

Area of Science:

• Environmental microbiology
• Computer vision
• Image analysis

Background:

• Scalable water-environment monitoring relies on environmental microorganism (EM) detection from microscopy images.
• Fair comparison of EM detection algorithms is difficult due to inconsistent protocols and limited benchmarks.

Purpose of the Study:

• Establish a reproducible benchmark for EM object detection using the EMDS-7 dataset.
• Evaluate 25 diverse object detection algorithms under unified conditions.
• Provide baseline performance metrics and diagnostic insights for EM detection research.

Main Methods:

• Constructed a benchmark on EMDS-7 with fixed data splits and unified input normalization.
• Trained 25 detectors from scratch, including two-stage, one-stage, keypoint-based, and Transformer-based methods.
• Evaluated performance using COCO-style mAP, AP across IoU thresholds, and recall analysis, with backbone comparisons (ResNet-18/50/101).

Main Results:

• Two-stage detectors (Faster R-CNN, Cascade R-CNN) achieved the highest accuracy (mAP 64.0%, 63.9%).
• Modern one-stage detectors (RTMDet-X) significantly closed the performance gap (mAP 60.9%).
• Performance decreased with stricter IoU thresholds, highlighting localization as a bottleneck; backbone scaling showed varied results.

Conclusions:

• The EMDS-7 benchmark favors detectors with robust localization and false positive control in cluttered microscopy images.
• Future research should prioritize high-IoU localization for small instances, boundary-aware learning, and false-positive suppression.
• The benchmark provides reproducible baselines to guide future environmental microorganism detection algorithm development.