TABLE 2.
Exploration of the performance of the FCOS in the grid search (IoU = 0.5)a
| Parameter | Result for a τ value of: |
|||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 0.05 | 0.1 | 0.15 | 0.2 | 0.25b | 0.3 | 0.35 | 0.4 | 0.5 | 0.6 | |
| Total number of ground truth objects | 2,387 | 2,479 | 2,526 | 2,552 | 2,563 | 2,564 | 2,568 | 2,572 | 2,575 | 2,575 |
| Blind spots | ||||||||||
| Number | 252 | 270 | 281 | 289 | 301 | 327 | 344 | 373 | 440 | 557 |
| Rate | 0.106 | 0.109 | 0.111 | 0.113 | 0.117 | 0.127 | 0.134 | 0.145 | 0.171 | 0.22 |
| Hallucinations | ||||||||||
| Number | 482 | 445 | 414 | 384 | 358 | 342 | 325 | 306 | 272 | 218 |
| Rate | 0.202 | 0.18 | 0.164 | 0.15 | 0.14 | 0.133 | 0.127 | 0.119 | 0.106 | 0.09 |
| Adjusted hallucinations | ||||||||||
| Number of true hallucinations | 8 | 10 | 13 | 14 | 17 | |||||
| Number misclassified | 22 | 26 | 30 | 33 | 37 | |||||
| Number correct | 384 | 348 | 315 | 295 | 271 | |||||
| Classifications | ||||||||||
| Number of errors | 389 | 387 | 384 | 378 | 352 | 332 | 319 | 299 | 261 | 215 |
| Number correct | 1,338 | 1,397 | 1,415 | 1,416 | 1,407 | 1,386 | 1,368 | 1,330 | 1,277 | 1,270 |
| Rate | 0.775 | 0.783 | 0.787 | 0.789 | 0.8 | 0.807 | 0.811 | |||
| Adjusted classifications | ||||||||||
| Number of errors | 406 | 404 | 382 | 365 | 356 | |||||
| Number correct | 1,799 | 1,764 | 1,722 | 1,681 | 1,639 | |||||
| Adjusted rate | 0.816 | 0.814 | 0.818 | 0.822 | 0.822 | |||||
| Sensitivity/recall | 0.842 | 0.838 | 0.834 | 0.83 | 0.824 | 0.809 | 0.799 | 0.781 | 0.736 | 0.65 |
| Adjusted sensitivity/recall | 0.865 | 0.859 | 0.851 | 0.837 | 0.827 | |||||
| Precision | 0.606 | 0.627 | 0.639 | 0.65 | 0.665 | 0.673 | 0.68 | 0.687 | 0.697 | 0.71 |
| Adjusted precision | 0.808 | 0.804 | 0.807 | 0.809 | 0.807 | |||||
| F 1 | 0.704 | 0.717 | 0.724 | 0.729 | 0.737 | 0.735 | 0.735 | 0.731 | 0.716 | 0.68 |
| Adjusted F1 | 0.838 | 0.834 | 0.832 | 0.826 | 0.821 | |||||
Relevant summary statistics for false negatives (blind spots), false-positive object predictions (hallucinations), and classifications. With a posterior analysis and correction of the hallucinations, performance was recalculated and reported in the adjusted sensitivity row. True positives correspond to a correctly identified object that is also correctly classified. In an FCOS, all pixels which are not part of a bounding box and which are not predicted to be part of a bounding box are true negatives. False positives are hallucinations and incorrect classifications. False negatives are blind spots only.
Boldface indicates estimates of performance for the optimal value of τ = 0.25; underlining indicates the final performance after adjustments.