Table 2.
Model comparison in terms of recall.
| Algorithm | AR1 | AR10 | ARM | ARL |
|---|---|---|---|---|
| Faster R-CNN | 0.30 ± 0.01 | 0.33 ± 0.01 | 0.07 ± 0.04 | 0.36 ± 0.02 |
| FCOS | 0.24 ± 0.02 | 0.24 ± 0.02 | 0.01 ± 0.01 | 0.27 ± 0.01 |
| RetinaNet | 0.31 ± 0.01 | 0.32 ± 0.01 | 0.06 ± 0.04 | 0.36 ± 0.02 |
| EfficientDet | 0.29 ± 0.01 | 0.30 ± 0.01 | 0.06 ± 0.02 | 0.34 ± 0.02 |
| YOLOv5 | 0.29 ± 0.01 | 0.30 ± 0.01 | 0.08 ± 0.03 | 0.34 ± 0.01 |
| NMS | 0.31 ± 0.01 | 0.32 ± 0.01 | 0.08 ± 0.04 | 0.36 ± 0.02 |
| NMW | 0.32 ± 0.01 | 0.33 ± 0.01 | 0.08 ± 0.04 | 0.37 ± 0.01 |
| Soft NMS | 0.31 ± 0.01 | 0.35 ± 0.01 | 0.08 ± 0.04 | 0.39 ± 0.02 |
| Soft Linear | 0.31 ± 0.01 | 0.36 ± 0.01 | 0.08 ± 0.04 | 0.40 ± 0.02 |
| WBF | 0.31 ± 0.01 | 0.33 ± 0.01 | 0.08 ± 0.04 | 0.37 ± 0.01 |
| WBF Max | 0.32 ± 0.01 | 0.33 ± 0.01 | 0.08 ± 0.04 | 0.37 ± 0.01 |
| Cascade R-CNN | 0.65 ± 0.02 | 0.69 ± 0.03 | 0.34 ± 0.12 | 0.75 ± 0.02 |
| StackBox with LR | 0.65 ± 0.03 | 0.71 ± 0.03 | 0.34 ± 0.03 | 0.76 ± 0.04 |
| StackBox with Adaboost | 0.42 ± 0.08 | 0.44 ± 0.09 | 0.08 ± 0.04 | 0.49 ± 0.10 |
| StackBox with RF | 0.65 ± 0.03 | 0.70 ± 0.03 | 0.34 ± 0.02 | 0.75 ± 0.04 |
| StackBox with GB | 0.64 ± 0.03 | 0.69 ± 0.03 | 0.31 ± 0.03 | 0.74 ± 0.04 |
| StackBox with XGBoost | 0.65 ± 0.03 | 0.70 ± 0.03 | 0.33 ± 0.03 | 0.75 ± 0.04 |
The results present the average values obtained by combining the 5 folds ± SD of those results. AR1 measures the average recall considering up to one detection per image, averaged over all IoUs, whereas AR10 considers 10 detections at most. Similar to precision, ARM measures the average recall on medium-sized ground-truth objects, whereas ARL only evaluates large ground-truth objects38. Bold denotes the highest values for each metric. The StackBox with Logistic Regression stands as the best model for all the metrics under consideration.