cs.CVNov 9, 2026

Correctness Metric Matters: NWD-Based Calibration Loss for Tiny Person Detection in UAV Infrared Imagery

Authors: Tian Zhang, Zhirong Su

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 LaECENWD_{\mathrm{NWD}} relative to BPC+IoU on all three groups (p<10−7p<10^{-7}) 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

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 23, 2026

The Paired Thermal Channel Is Two Orthogonal Axes: A Controlled Attribution Study of Calibration Cost and Detection Gain in Infrared Person Detection

The calibrated thermal channel in infrared person detection comprises two structurally orthogonal axes. True calibrated ΔTΔT outperforms scale-matched grayscale pseudo-ΔTΔT by +0.0221+0.0221 mAP (paired Wilcoxon p=0.0098p=0.0098, Cohen's d=+1.08d=+1.08), confirming that the calibrated channel carries information the grayscale channel lacks. The calibration axis is driven by the joint pairing-scale condition. Preserving both paired correspondence and absolute calibrated scale raises expected calibration error (ECE) from ≈0.05\approx 0.05 to 0.1420.142. The 2×22\times 2 interaction (p=0.0059p=0.0059) localizes the rise to this condition. Breaking either factor alone returns ECE to the broken-factor level (true-vs-const p=0.002p=0.002; true-vs-shuffled p=0.002p=0.002). Under post-hoc temperature scaling, the pairing cost survives recalibration (p=0.002p=0.002), while the scale cost is largely absorbed (p=0.16p=0.16). As regards the detection axis, the gain stems from structural de-structuring. Phase randomization, which destroys spatial structure while preserving per-sample variance, yields +0.0185+0.0185 mAP over true ΔTΔT on FLIR ADAS (p=0.014p=0.014, d=+0.96d=+0.96). The FLAME3 gain is +0.0064+0.0064 mAP (p=0.078p=0.078, n=8). On the calibration axis, the same intervention is insignificant on FLIR ADAS (p=0.47p=0.47) but significant on FLAME3 (p=0.0078p=0.0078). These findings rest on a five-arm controlled attribution with temperature-evidential learning (TEL) as the testbed. TEL obtains a single evidence-derived routing scalar together with a complementary diagnostic entropy, and establishes a finite-evidence degeneracy theorem proving that the acquisition product degenerates to this scalar with a closed-form bound. A physical 0.5∘0.5^\circC threshold admits cross-dataset physical readings, whereas a gray threshold of 2020 inverts without physical interpretation.
Apr 11, 2026

Revisiting the Scale Loss Function and Gaussian-Shape Convolution for Infrared Small Target Detection

Infrared small target detection still faces two persistent challenges: training instability from non-monotonic scale loss functions, and inadequate spatial attention due to generic convolution kernels that ignore the physical imaging characteristics of small targets. In this paper, we revisit both aspects. For the loss side, we propose a diff-based scale loss that weights predictions according to the signed area difference between the predicted mask and the ground truth, yielding strictly monotonic gradients and stable convergence. We further analyze a family of four scale loss variants to understand how their geometric properties affect detection behavior. For the spatial side, we introduce Gaussian-shaped convolution with a learnable scale parameter to match the center-concentrated intensity profile of infrared small targets, and augment it with a rotated pinwheel mask that adaptively aligns the kernel with target orientation via a straight-through estimator. Extensive experiments on IRSTD-1k, NUDT-SIRST, and SIRST-UAVB demonstrate consistent improvements in mIoUmIoU, PdP_d, and FaF_a over state-of-the-art methods. We release our anonymous code and pretrained models.
Apr 14, 2026

DroneScan-YOLO: Redundancy-Aware Lightweight Detection for Tiny Objects in UAV Imagery

Aerial object detection in UAV imagery presents unique challenges due to the high prevalence of tiny objects, adverse environmental conditions, and strict computational constraints. Standard YOLO-based detectors fail to address these jointly: their minimum detection stride of 8 pixels renders sub-32px objects nearly undetectable, their CIoU loss produces zero gradients for non-overlapping tiny boxes, and their architectures contain significant filter redundancy. We propose DroneScan-YOLO, a holistic system contribution that addresses these limitations through four coordinated design choices: (1) increased input resolution of 1280x1280 to maximize spatial detail for tiny objects, (2) RPA-Block, a dynamic filter pruning mechanism based on lazy cosine-similarity updates with a 10-epoch warm-up period, (3) MSFD, a lightweight P2 detection branch at stride 4 adding only 114,592 parameters (+1.1%), and (4) SAL-NWD, a hybrid loss combining Normalized Wasserstein Distance with size-adaptive CIoU weighting, integrated into YOLOv8's TaskAligned assignment pipeline. Evaluated on VisDrone2019-DET, DroneScan-YOLO achieves 55.3% mAP@50 and 35.6% mAP@50-95, outperforming the YOLOv8s baseline by +16.6 and +12.3 points respectively, improving recall from 0.374 to 0.518, and maintaining 96.7 FPS inference speed with only +4.1% parameters. Gains are most pronounced on tiny object classes: bicycle AP@50 improves from 0.114 to 0.328 (+187%), and awning-tricycle from 0.156 to 0.237 (+52%).