Correctness Metric Matters: NWD-Based Calibration Loss for Tiny Person Detection in UAV Infrared Imagery
Organizations: Shandong Management University, Jinan 250357, China. · Tsinghua University, Beijing 100084, China
Abstract
Detection calibration aligns predicted confidence with empirical correctness. Existing calibration losses inherit their correctness definition from box regression, specifically IoU against a hard threshold. We show that this target is geometrically ill-posed for tiny objects in UAV infrared imagery, where persons occupy less than 0.1% of the image. The same pixel error is penalized more heavily as the target shrinks, so binary prediction calibration (BPC) actively suppresses confidences that should be high. On HIT-UAV, BPC with an IoU target degrades image-level localization-aware ECE (LaECE) from 0.0689 to 0.0761, 0.0649 to 0.0733, and 0.0653 to 0.0682 across Low, Mid, and High signal-to-clutter ratio (SCR) groups. We replace the IoU target with a continuous, scale-invariant normalized Wasserstein distance (NWD). The NWD target reduces LaECE relative to BPC+IoU on all three groups () with mAP neutral (0.9356 vs.\ 0.9347), and an independent center-offset anchor confirms that the gain is a genuine geometric correction. The improvement holds across a broad plateau of calibration weights. The NWD formulation requires no inference-time mapping layer. If a validation-fitted isotonic map is applied on top of it, calibration is lowered further, and the two corrections act on different components of the same error. The metric-level remedy has a measurable boundary. Target-background thermal pairing in the training distribution and detection-level aggregation under per-image detection-count heterogeneity lie outside it. Both are quantified independently of the NWD target.
Figures & tables
| Group | LaECE NWD | LaECE IoU | ||
|---|---|---|---|---|
| Low | ||||
| Mid | ||||
| High | ||||
| Sum | ||||
| Condition | ||||
|---|---|---|---|---|
| No BPC (50 ep) | ||||
| BPC+IoU ( ) | ||||
| BPC+NWD ( ) | ||||
| + Temperature Scaling | ||||
| + Fair Isotonic |
| A | B | (A B) | |
|---|---|---|---|
| mAP 50 | 0.9284 | 0.9253 | |
| mAP | 0.5134 | 0.5094 | |
| LaECE Low | 0.0726 | 0.0494 | ( ) |
| LaECE Mid | 0.0627 | 0.0492 | ( ) |
| LaECE High | 0.0617 | 0.0442 | ( ) |
| Block | LaECE | LaECE | mAP | |||||
|---|---|---|---|---|---|---|---|---|
| Low | Mid | High | Low | Mid | High | 50 | ||
| Group | LaECE (image) | D-ECE (det.) | D-ECE (image proxy) |
|---|---|---|---|
| Low | 0.0728 | 0.0385 | 0.1185 |
| Mid | 0.0627 | 0.0701 | 0.1166 |
| High | 0.0617 | 0.0739 | 0.1158 |
| Condition | mAP | LaECE NWD (L/M/H) | LaECE IoU (L/M/H) | |||
|---|---|---|---|---|---|---|
| 50 | Mean | Std | Mean | Std | ||
| No BPC (50 ep) | ||||||
| BPC+IoU | ||||||
| BPC+NWD (Ours) | ||||||
| Condition | px | px (LaECE) | |||||
|---|---|---|---|---|---|---|---|
| Low | Mid | High | Brier | Low | Mid | High | |
| No BPC | 0.1054 | 0.1233 | 0.1322 | 0.0419 | 0.1074 | 0.1252 | 0.1323 |
| BPC+IoU | 0.1204 | 0.1408 | 0.1384 | 0.0455 | 0.1211 | 0.1422 | 0.1382 |
| BPC+NWD (Ours) | 0.1113 | 0.1262 | 0.1309 | 0.0415 | 0.1109 | 0.1275 | 0.1309 |
| LaECE | LaECE | |||||
|---|---|---|---|---|---|---|
| Low | Mid | High | Low | Mid | High | |
| 20 | ||||||
| 35 | ||||||
| 50 | ||||||
| 65 | ||||||
| 100 | ||||||
| LaECE (L/M/H) | LaECE (L/M/H) | mAP 50 | mAP | |
|---|---|---|---|---|
| 10 | 0.0815/0.0773/0.0740 | 0.0989/0.0996/0.0970 | 0.9364 | 0.5133 |
| 15.6 | 0.0725/0.0718/0.0726 | 0.0868/0.0916/0.0915 | 0.9367 | 0.5156 |
| 20 | 0.0718/0.0673/0.0692 | 0.0858/0.0847/0.0871 | 0.9316 | 0.5133 |
| 25 | 0.0763/0.0705/0.0687 | 0.0915/0.0887/0.0853 | 0.9356 | 0.5107 |
| 30 | 0.0807/0.0746/0.0724 | 0.0959/0.0938/0.0900 | 0.9295 | 0.5081 |
| mAP | LaECE | LaECE | ||||||
|---|---|---|---|---|---|---|---|---|
| 50 | Low | Mid | High | Low | Mid | High | ||
| A: IoU match + NWD target | 0.0861 | 0.0914 | 0.0868 | 0.0711 | 0.0715 | 0.0671 | ||
| B: NWD match + NWD target | 0.0854 | 0.0804 | 0.0686 | 0.0911 | 0.0972 | 0.0928 | ||
| Condition | LaECE NWD | LaECE IoU | mAP 50 | ||||
|---|---|---|---|---|---|---|---|
| Low | Mid | High | Low | Mid | High | ||
| IoU Baseline | 0.9366 | ||||||
| NWD Low-IoU Only | 0.9345 | ||||||
| NWD High-IoU Only | 0.9337 | ||||||
| SizeAdapt ( ) | |||||||
| SizeAdapt ( ) | |||||||
| Method | LaECE IoU (independent anchor) | LaECE NWD (aligned anchor) | ||||
|---|---|---|---|---|---|---|
| Low | Mid | High | Low | Mid | High | |
| BPC+IoU (no post-proc.) | 0.0761 | 0.0733 | 0.0682 | 0.0957 | 0.0993 | 0.0948 |
| + Temperature Scaling | 0.0718 | 0.0765 | 0.0831 | 0.0683 | 0.0583 | 0.0556 |
| + Fair Post-hoc (Iso on Val) | 0.0671 | 0.0635 | 0.0611 | 0.0874 | 0.0891 | 0.0868 |
| + Oracle Isotonic (test-fit) | 0.0661 | 0.0623 | 0.0602 | 0.0867 | 0.0870 | 0.0845 |
| + NWD (Ours, end-to-end) | 0.0711 | 0.0715 | 0.0671 | 0.0861 | 0.0914 | 0.0868 |
| Condition | LaECE NWD | mAP | |||
|---|---|---|---|---|---|
| Low | Mid | High | 50 | ||
| BPC+IoU | 0.9366 | 0.5169 | |||
| BPC+NWD (Ours) | 0.9367 | 0.5156 | |||
| NWD-Matched | 0.9287 | 0.5107 | |||
| SCR | Size (px 2 ) | BPC+IoU | BPC+NWD | |
|---|---|---|---|---|
| (NWD IoU) | ||||
| High | 0.1137 | 0.0988 | ||
| – | 0.0870 | 0.1021 | ||
| 0.1343 | 0.1335 | |||
| Mid | 0.1348 | 0.1228 | ||
| – | 0.1055 | 0.0990 |
| Subset | BPC+IoU | BPC+NWD | (NWD IoU) |
|---|---|---|---|
| IoU 0.5 | 0.0634 | 0.0695 | |
| IoU 0.5 | 0.1915 | 0.1554 |
| mAP | LaECE NWD | LaECE IoU | ||||||
| Condition | 50 | Low | Mid | High | Low | Mid | High | |
| No BPC | 0.7405 | 0.4373 | 0.0886 | 0.0786 | 0.0836 | 0.0705 | 0.0620 | 0.0657 |
| BPC+IoU | 0.7326 | 0.4311 | 0.1005 | 0.0774 | 0.0798 | 0.0816 | 0.0610 | 0.0622 |
| BPC+NWD | 0.7225 | 0.4263 | 0.0753 | 0.0784 | 0.0796 | 0.0637 | 0.0633 | 0.0628 |
| Group | LaECE (image) | D-ECE (det.) |
|---|---|---|
| High | 0.0692 | 0.0504 |
| Mid | 0.0658 | 0.0558 |
| Low | 0.0580 | 0.0648 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Condition | ||||
|---|---|---|---|---|
| 20 | BPC+IoU | |||
| BPC+NWD | ||||
| 35 | BPC+IoU | |||
| BPC+NWD | ||||
| 50 | BPC+IoU | |||
| BPC+NWD |
| Group | NWD+Iso | IoU+Iso | |||
|---|---|---|---|---|---|
| seed 0 | 533 | 0.0188 | 0.0175 | 0.0127 | |
| seed 1 | 537 | 0.0177 | 0.0155 | 0.685 | |
| seed 42 | 512 | 0.0184 | 0.0186 | 0.0026 | |
| 3-seed combined | 490 | 0.0197 | 0.0185 | 0.032 | |
| 88 | 0.0255 | 0.0160 | 0.092 | ||
| 6–20 | 216 | 0.0175 | 0.0184 | 0.0098 |
Object Detection
Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.
See all 671 papers