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

Tian Zhang; Yang Chen; Bing Bai; Weikai He; Xiang Gao; Zhirong Su

cs.CVOct 23, 2026

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

Authors: Tian Zhang, Yang Chen, Bing Bai, Weikai He, Xiang Gao, Zhirong Su

Organizations: School of Information Engineering, Shandong Management University, Jinan, China. · School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore. · School of Aeronautics, Shandong Jiaotong University, Jinan, China. · School of Business, Shandong Management University, Jinan, China.

Abstract

The calibrated thermal channel in infrared person detection comprises two structurally orthogonal axes. True calibrated $ΔT$ outperforms scale-matched grayscale pseudo- $ΔT$ by $+0.0221$ mAP (paired Wilcoxon $p=0.0098$ , Cohen's $d=+1.08$ ), 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 $\approx 0.05$ to $0.142$ . The $2\times 2$ interaction ( $p=0.0059$ ) localizes the rise to this condition. Breaking either factor alone returns ECE to the broken-factor level (true-vs-const $p=0.002$ ; true-vs-shuffled $p=0.002$ ). Under post-hoc temperature scaling, the pairing cost survives recalibration ( $p=0.002$ ), while the scale cost is largely absorbed ( $p=0.16$ ). As regards the detection axis, the gain stems from structural de-structuring. Phase randomization, which destroys spatial structure while preserving per-sample variance, yields $+0.0185$ mAP over true $ΔT$ on FLIR ADAS ( $p=0.014$ , $d=+0.96$ ). The FLAME~~3 gain is $+0.0064$ mAP ( $p=0.078$ , n=8). On the calibration axis, the same intervention is insignificant on FLIR ADAS ( $p=0.47$ ) but significant on FLAME~~3 ( $p=0.0078$ ). 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 $0.5^\circ$ C threshold admits cross-dataset physical readings, whereas a gray threshold of $20$ inverts without physical interpretation.

Figures & tables

Figure 1: Temperature-evidential learning (TEL) as one evidence chain. (a) Dual-channel decoding. Channel A renders

G

, while Channel B retains the calibrated temperature matrix

T

. The annular contrast geometry is sampled at dilation factors

r\in\{0.1,0.2,0.4\}

, using

T

values from Channel B and box positions from the detector backbone. (b) Temperature-evidential modulation. The multi-scale contrast

\Delta T^{(s)}

(Eq. ( 3 )) passes through the sigmoid gate

m(\Delta T)

(Eq. ( 4 )). The temperature-aware prior

g_{k},\pi_{k}

produces

e_{\text{aug},k}

(Eq. ( 9 )), and the modulation factor

1+\gamma m

produces

e_{\text{mod},k}

(Eq. ( 10 )), from which a single evidence-derived routing scalar

u

is read; a diagnostic entropy

\Psi=\mathbb{E}[H(p)]

is computed alongside on a dashed branch and does not participate in routing. (c) Two-ended routing. The same

u

drives annotation, pseudo-labeling, and discard (Eq. ( 22 )); the partition relies on separation of the two ends of the distribution, not on monotone ranking in between. (d) Replay and distillation weighting (Eqs. ( 35 ) and ( 37 )) appear as dashed design extensions. Solid arrows denote forward data flow and core computation. Dashed arrows denote auxiliary inputs, control signals, and the closed-loop feedback that returns the updated detector to the next round. All downstream decisions in (b)-(d) use the temperature matrix retained in Channel B.

Panel of Figure 2 — Figure 2: Physical evidence of the temperature matrix. (a) Two nighttime patches in which the grayscale contrast is low ( $|\Delta G|\approx 2.2$ ) while the calibrated $\Delta T$ remains measurable ( $\approx 0.28^{\circ}$ C). The inset text reports the exact number of interior and annular background pixels. (b) Conditional distribution $P(\Delta T\mid y{=}\text{person})$ as a function of the illumination drift coordinate $s$ , stratified into quintiles. $P(y\mid s)$ is stationary, while $P(\Delta T\mid y,s)$ drifts with $s$ . Data points mark median $\pm$ IQR. (c) Empirical validation of the KL separability sandwich in Eq. ( 19 ). The red line is the measured median $\Lambda(b)$ with interquartile range (IQR) error bars, obtained by using the thermal camera’s native 8-bit grayscale rendering (strictly pixel-aligned with the 14-bit TIFF) to compute the true local grayscale ratio $x$ . The black dashed line is the theoretical upper bound $\overline{\Lambda}(b)$ . The empirical median grayscale-to-temperature noise ratio is $x\approx 1.10$ . The empirical median strictly falls within the theoretical bounds, confirming that the calibrated temperature channel $\Delta T$ provides a reliable source of local separability evidence. The empirical median temperature noise is $\sigma_{med}=0.39^{\circ}$ C (IQR: $[0.22,0.72]^{\circ}$ C).

Signal	$\varepsilon$	$\rho$	$\text{err}\|\text{acc}$	ub	Coverage
Mean $\pm$ std across the $45$ seed pairs
$u$ / margin / conf	$0.086\pm 0.021$	$0.712\pm 0.072$	$0.0364\pm 0.014$	$0.0660\pm 0.016$	$45/45$
$H$	$0.116\pm 0.021$	$0.520\pm 0.084$	$0.0281\pm 0.010$	$0.0736\pm 0.014$	$45/45$
$95\%$ bootstrap CI (2000 resamples, pair as unit)
$u$ / margin / conf	$[0.080,0.092]$	$[0.690,0.734]$	$[0.032,0.041]$	$[0.061,0.071]$	—
$H$	$[0.110,0.122]$	$[0.495,0.545]$	$[0.025,0.031]$	$[0.070,0.078]$	—

Table 1: Empirical validation of Eq. ( 27 ) over all

45

seed pairs. Margin and conf are identical to each other;

u

is numerically indistinguishable from them because

K=2

. Coverage is the fraction of seed pairs for which the bound exceeds the accepted error.

Signal-to-signal correlations
Pair	Pearson $\rho$	95% CI	Spearman $\rho$	95% CI
$u$ , $H$	$-0.0765$	$[-0.0782,-0.0748]$	$-0.2662$	$[-0.2687,-0.2639]$
$u$ , margin	$-0.9255$	$[-0.9263,-0.9247]$	$-0.3079$	$[-0.3109,-0.3047]$
$u$ , conf	$-0.9255$	$[-0.9263,-0.9247]$	$-0.3079$	$[-0.3110,-0.3050]$
$H$ , margin	$-0.2590$	$[-0.2608,-0.2573]$	$-0.6202$	$[-0.6224,-0.6180]$
$H$ , conf	$-0.2590$	$[-0.2608,-0.2573]$	$-0.6202$	$[-0.6225,-0.6180]$

Table 2: Signal correlations on the FLIR ADAS prediction protocol, pooled over ten seeds. Values are Pearson

\rho

and Spearman

\rho

with

95\%

bootstrap confidence intervals on the pooled RoIs.

$\tau_{\text{low}}$	measured (mean $\pm$ std)	$[\min,\max]$	bound
$0.40$	$0.0802\pm 0.0704$	$[0.0126,0.2231]$	$0.1667$
$0.45$	$0.0308\pm 0.0158$	$[0.0129,0.0592]$	$0.1837$
$0.50$	$0.0237\pm 0.0058$	$[0.0162,0.0378]$	$0.2000$
$0.55$	$0.0332\pm 0.0051$	$[0.0254,0.0454]$	$0.2157$
$0.60$	$0.0470\pm 0.0027$	$[0.0423,0.0524]$	$0.2308$

Table 3: Empirical validation of Corollary 3 on an oracle-based image-level protocol, averaged over ten seeds. Thresholds below

0.40

are omitted because the discard branch contains too few samples for a reliable estimate.

$a$	$b$	consistency $\kappa_{\text{route}}$	$(q_{\text{lo}},q_{\text{hi}})$
$1.0$	$0.0$	$1.0000$	$(0.4520,0.9883)$
$0.5$	$0.0$	$1.0000$	$(0.4520,0.9883)$
$1.0$	$10.0$	$1.0000$	$(0.4520,0.9883)$
$2.0$	$-5.0$	$1.0000$	$(0.4520,0.9883)$

Table 4: Empirical validation of Corollary 4 . Four strictly increasing affine reparametrizations of the temperature channel,

T^{\prime}=aT+b

, are applied; the routing partition is unchanged in all four cases.

	mean	std	var
$u$	0.3673	0.2175	0.047301
$\Psi$	0.4982	0.0204	0.000416
$\Delta$	0.1827	0.1086	0.011792

Table 5: Empirical statistics of the routing scalar, the diagnostic entropy, and the acquisition product, pooled over

n=5.4\times 10^{5}

RoIs and ten seeds on FLIR ADAS.

regime	$u$ range	$\mathrm{frac}(\mathrm{IoU}>0.5)$	mean IoU	router branch
background	bins 0–6	$\approx 0$	$\approx 0$	discard
transition	bin 7	$0.0948$	$0.0861$	pseudo
detection	bin 8	$0.9274$	$0.7652$	pseudo
saturated	bin 9	$0.3184$	$0.2895$	annotate

Table 6: Three-state structure of the routing scalar

u

on FLIR ADAS, pooled over ten seeds. The

u_{\mathrm{med}}

values of the four regimes are

0.263

–

0.271

(bins 0–6),

0.2718

(bin 7),

0.5620

(bin 8), and

0.9943

(bin 9). Regime labels describe the

u

-stratified structure. Router branches follow Eq. ( 23 ) with pooled ten-seed thresholds

Q_{0.3}=0.267

and

Q_{0.9}=0.714

. Bins 0–6 are grouped as a single background regime. Their median

u

spans

Q_{0.3}

, so some samples within this regime route to pseudo. The saturated regime is the only regime where mean IoU decreases relative to the preceding bin; this drop is what makes

u

a two-ended separator and not a monotone ranking.

Dataset	Modality	Calibrated $T$ ?	Role
FLIR ADAS	RGB + 14-bit TIFF	✓	Primary
FLAME 3	Radiometric Long-Wave Infrared (LWIR)	✓	Auxiliary

Table 7: Datasets used in this work.

Contrast	Metric	$\Delta$	$t$ -test $p$	Wilcoxon $p$	Cohen’s $d$	Reading
true $-$ pseudo	mAP	$+0.0221$	$0.0077$	$\mathbf{0.0098}$ ( $W=3$ )	$+1.08$	primary, Wilcoxon
true $-$ shuffled	ECE	$+0.0882$	$0.039$	$\mathbf{0.002}$ ( $W=0$ )	$+0.76$	primary, true higher
true $-$ pseudo	ECE	$-0.0097$	$0.77$	$0.064$	$-0.095$	secondary, n.s.
true $-$ pseudo	Recall	$+0.0008$	$0.66$	n.s.	$+0.14$	secondary, n.s.
true $-$ pseudo	LCR	$+0.0044$	$0.60$	n.s.	$+0.17$	secondary, n.s.
true $-$ const- $\Delta T$	ECE	$+0.0916$	$0.019$	$\mathbf{0.002}$ ( $W=0$ )	$+0.90$	secondary, Holm-pass

Table 8: Paired contrasts on FLIR ADAS over ten seeds under the two-stage protocol. All Cohen’s

d

values are paired-effect sizes, computed from the per-seed paired differences. The primary family is true-vs-pseudo on mAP and true-vs-shuffled on ECE (

m=2

, Holm threshold

0.025

); the secondary family is true-vs-pseudo ECE/Recall/LCR and true-vs-const-

\Delta T

ECE (

m=4

, Holm thresholds

0.0125

0.0167

0.025

). The detection-axis attribution (var-phase-vs-true mAP and ECE) is pre-specified and reported as a dedicated contrast pair. Rows below the third rule are exploratory. Positive

\Delta

means the first arm is larger.

Probe	$e_{\text{mod}}$ mean	$\Delta e_{\text{mod}}$	Top score
zero modulation	$0.002453$	reference	$0.8647$
true $\Delta T$	$0.003395$	$+38.42\%$ ( $p=4.2\times 10^{-25}$ )	$0.8862$ ( $+2.48\%$ , $p=8.7\times 10^{-43}$ )
random modulation	$0.002668$	$+8.78\%$ ( $p=2.97\times 10^{-2}$ )	$0.8651$ ( $+0.05\%$ , $p=9.5\times 10^{-3}$ )

Table 9: Attribution-consumption probe on FLIR ADAS (

200

images). The modulated evidence magnitude

e_{\text{mod}}

and the top detection score are compared across zero modulation, true calibrated

\Delta T

, and random modulation.

Variant	[email protected]	Recall	LCR	ECE $\downarrow$	Params/FLOPs matched?
Baseline (no modulation)	$0.4664\pm 0.0000$	$0.3761\pm 0.0000$	$0.4279\pm 0.0000$	$0.4585\pm 0.0000$	–
Random modulation	$0.5190\pm 0.0096$	$0.3767\pm 0.0028$	$0.4247\pm 0.0117$	$0.0373\pm 0.0068$	✓
Grayscale pseudo- $\Delta T$ (matched)	$0.4628\pm 0.0135$	$0.3750\pm 0.0040$	$0.4211\pm 0.0067$	$0.1512\pm 0.0298$	✓
Shuffled true $\Delta T$	$0.4804\pm 0.0105$	$0.3744\pm 0.0040$	$0.4228\pm 0.0148$	$0.0534\pm 0.0135$	✓
const- $\Delta T$	$0.4833\pm 0.0140$	$0.3789\pm 0.0040$	$0.4240\pm 0.0142$	$0.0500\pm 0.0210$	✓
TEL (true $\Delta T$ )	$0.4849\pm 0.0152$	$0.3758\pm 0.0024$	$0.4255\pm 0.0136$	$0.1415\pm 0.1069$	✓

Table 10: Overall comparison on FLIR ADAS (10-seed prediction, all bins pooled). The primary acceptance test is TEL(true

\Delta T

) versus Grayscale pseudo-

\Delta T

true $-$ pseudo $=+0.0041$ mAP (paired Wilcoxon $p=0.082$ ); true $-$ random $=-0.0077$ ( $p=0.0074$ );
Variant	[email protected]	Recall	Low-contrast Recall	ECE $\downarrow$
Random modulation	$0.9238\pm 0.0053$	$0.9023\pm 0.0018$	$0.9337\pm 0.0048$	$0.0421\pm 0.0056$
Grayscale pseudo- $\Delta T$	$0.9120\pm 0.0060$	$0.8993\pm 0.0013$	$0.9281\pm 0.0037$	$0.3066\pm 0.0124$
TEL (true $\Delta T$ )	$0.9161\pm 0.0038$	$0.9005\pm 0.0025$	$0.9300\pm 0.0064$	$0.3095\pm 0.0230$
const- $\Delta T$	$0.9153\pm 0.0074$	$0.8988\pm 0.0025$	$0.9281\pm 0.0073$	$0.2319\pm 0.0662$
var-phase	$0.9225\pm 0.0056$	$0.9019\pm 0.0025$	$0.9342\pm 0.0064$	$0.0639\pm 0.0058$

Table 11: Auxiliary validation on FLAME 3 (radiometric wildfire detection,

|\Delta T|\approx 30

–

60^{\circ}

C). Mean

\pm

std over eight seeds. All

p

values are paired Wilcoxon signed-rank tests over the eight paired seeds, where the smallest attainable two-sided

p

0.0078

Variant	ECE (raw)	$T$ (median)	ECE ${}_{\text{TS}}$	$\Delta$ ECE
Baseline (no modulation)	$0.4585$	$2.645$	$0.4588$	$+0.0003$
Random modulation	$0.0373$	$0.807$	$0.0166$	$-0.0206$
Grayscale pseudo- $\Delta T$	$0.1512$	$0.661$	$0.0860$	$-0.0652$
Shuffled true $\Delta T$	$0.0534$	$0.706$	$0.0172$	$-0.0361$
const- $\Delta T$	$0.0500$	$0.872$	$0.0342$	$-0.0156$
TEL (true $\Delta T$ )	$0.1415$	$0.763$	$0.0837$	$-0.0578$

Table 12: Post-hoc temperature scaling on FLIR ADAS (10 seeds).

T

is fitted per arm on the validation set with NLL as the objective;

\mathrm{ECE}_{\text{TS}}

uses

M=15

T

is the median across seeds; the true arm’s mean is affected by seed s2 (

T=3.89

; Section 4.5 ). Raw ECE is reproduced from Table 10 ;

\Delta

ECE is computed from per-seed means before rounding.

Variant	paired	absolute	mean	p50	p90
none	–	–	0.493	0.500	0.500
random	$\times$	$\times$	0.083	0.032	0.062
true	✓	✓	0.222	0.110	0.637
shuffled	$\times$	✓	0.115	0.062	0.164
const- $\Delta T$	✓	$\times$	0.158	0.059	0.652

Table 13: Confidence distribution over 10 seeds (paired protocol).

Quantity	Value	95% CI	$p$
Interaction contrast $\bar{c}$	$+0.0756$	$[+0.011,+0.140]$	$t$ -test $0.0475$ ; Wilcoxon $0.0059$
Cohen’s $d$	$+0.73$	—	—
Leave-one-out $\bar{c}$ (sans s2)	$+0.0437$	—	$<0.005$
Component $\mathrm{ECE}_{\text{true}}-\mathrm{ECE}_{\text{const}}$	$+0.0916$	—	—
Component $\mathrm{ECE}_{\text{shuffled}}-\mathrm{ECE}_{\text{random}}$	$+0.0161$	—	—

Table 14: Joint-condition interaction test on ECE over ten paired seeds. The 2

\times

2 contrast

c=(\mathrm{ECE}_{1,1}-\mathrm{ECE}_{1,0})-(\mathrm{ECE}_{0,1}-\mathrm{ECE}_{0,0})

is positive and significant under the Wilcoxon signed-rank test.

Figure 3:

\Delta

ECE (true

-

random) versus per-box separability ratio

\rho

bins, with GT-derived quintile boundaries

q_{20}/q_{40}/q_{60}/q_{80}=0.703/1.271/1.843/2.588

Figure 4: Cross-dataset reversal of the two axes. The same phase-randomized intervention (var-phase

-

true) produces a detection-axis effect of the same sign and comparable magnitude on FLIR ADAS and FLAME 3, and a calibration-axis effect that is absent on FLIR ADAS (

p=0.47

) and significant on FLAME 3 (

p=0.0078

, the smallest attainable at eight seeds). Each axis is significant on the dataset where its structural counterpart is most pronounced, and insignificant on the other.

Dataset	$\Delta T$ SNR	Pearson $r$	Discriminates
FLIR ADAS	channel SNR median $8.96$ (IQR $5.75$ – $13.98$ ); per-box sep. SNR $0.69$	low (collapse)	true vs. pseudo
FLAME 3	$\|\Delta T\|$ p25/p50/p75 $=29.89/42.04/59.58^{\circ}$ C	Pearson $r=0.76$ (redundant)	true vs. pseudo redundancy; random superiority check

Table 15: Data conditions and their discriminating power. Each dataset supplies a distinct validation axis.

Panel of Figure 5 — Figure 5: Cross-dataset physical contrast.

Threshold $\tau$	FLIR person pass	FLAME3 fire pass	Physical reading
$0.5^{\circ}$ C	$10.5\%$	$99.95\%$	FLIR: cold-background persons; FLAME 3: all fires
$5^{\circ}$ C	$0\%$	$99.9\%$	FLIR: no person reaches $5^{\circ}$ C; FLAME 3: most fires
$40^{\circ}$ C	$0\%$	$53.5\%$	FLIR: no person reaches $40^{\circ}$ C; FLAME 3: strong fires only
$\tau_{G}=20$ (8-bit gray)	$73.6\%$	$36.0\%$	no physical reading

Table 16: Cross-dataset pass rates of physical and gray thresholds.

Figure 6: [email protected] versus

\Delta T

bucket midpoint on FLIR ADAS for five variants: none, pseudo, random, TEL (true), and var-phase. All five curves rise with

\Delta T

up to B4, the high-transition bucket, and then decline at the redundant end (B5). The var-phase curve lies above true at B1, B2, and B4, below it at B3 and B5, with the largest positive gap at B4, the bucket where the injected-variance and true-variance scales are best matched. This is the qualitative counterpart of the paired

+0.0185

mAP effect reported in Table 8 . Buckets are test-set GT

\Delta T

median quintiles with boundaries

q_{20}/q_{40}/q_{60}/q_{80}=0.2424/0.2721/0.3392/0.5015^{\circ}

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Signal	Pair	$\text{err}\|\text{acc}$	ub	Gap
$u$	$(7,8)$	$0.0556$	$0.0575$	$+0.0019$
$u$	$(2,9)$	$0.0273$	$0.0352$	$+0.0078$
$u$	$(1,8)$	$0.0616$	$0.0698$	$+0.0082$
$u$	$(3,9)$	$0.0360$	$0.0480$	$+0.0120$
$u$	$(2,3)$	$0.0267$	$0.0396$	$+0.0129$
$H$	$(7,8)$	$0.0348$	$0.0551$	$+0.0203$

Appendix

Table 17: Tightest seed pairs for the pseudo-label error bound in Eq. ( 27 ). Only

u

and

H

are shown; margin and conf are omitted because they track

u

under

K=2

Seed	$\tau{=}0.40$	$\tau{=}0.45$	$\tau{=}0.50$	$\tau{=}0.55$	$\tau{=}0.60$
0	0.0140	0.0144	0.0227	0.0355	0.0486
1	0.0619	0.0234	0.0162	0.0340	0.0484
2	0.1667	0.0541	0.0285	0.0311	0.0463
3	0.0436	0.0306	0.0245	0.0311	0.0423
4	0.0179	0.0179	0.0192	0.0295	0.0464
5	0.2231	0.0592	0.0378	0.0454	0.0524

Appendix

Table 18: Per-seed discard error

\Pr(\hat{y}\neq y\mid u<\tau_{\text{low}})

under the oracle-based image-level protocol.

Pearson $\rho$
Pair	s0	s1	s2	s3	s4	s5	s6	s7	s8	s9
$u$ , $H$	$-0.092$	$-0.145$	$-0.008$	$-0.075$	$-0.009$	$-0.060$	$-0.013$	$-0.147$	$-0.062$	$-0.189$
$u$ , margin	$-0.922$	$-0.930$	$-0.943$	$-0.922$	$-0.921$	$-0.929$	$-0.932$	$-0.926$	$-0.915$	$-0.921$
$H$ , margin	$-0.259$	$-0.191$	$-0.275$	$-0.267$	$-0.330$	$-0.268$	$-0.298$	$-0.193$	$-0.296$	$-0.169$
margin, conf	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$	$+1.000$
$\Delta$ , $H$	$-0.027$	$-0.091$	$+0.047$	$-0.008$	$+0.057$	$-0.005$	$+0.049$	$-0.089$	$+0.007$	$-0.130$

Appendix

Table 19: Per-seed signal correlations on the FLIR ADAS prediction protocol. Redundant pairs (

u

–conf,

H

–conf,

\Delta

–conf) are omitted; they are identical to the corresponding margin pairs because

K=2

FLIR subset correlations
Set	$n$	Signed Pearson	Signed Spearman	Unsigned Pearson	Unsigned Spearman
FLIR train	48,649	$+0.7126$	$+0.7183$	$+0.5068$	$+0.5709$
FLIR test	12,267	$+0.8136$	$+0.8847$	$+0.8073$	$+0.8812$
FLAME 3	2,126	$+0.7570$	$+0.8450$	$+0.7569$	$+0.8450$
FLIR subset			Spearman		$n$
LCR ( $\|\Delta T\|<$ p25)			$+0.9338$		3,067

Appendix

Table 20:

\Delta G

–

\Delta T

correlation.

Dataset	Quantity	p25	p50	p75	Other
FLIR ADAS	$\|\Delta T\|$ ( ∘ C)	$0.208$	$0.271$	$0.357$	LOW_DT $0.207$
FLIR ADAS	eff. SNR	$5.75$	$8.96$	$13.98$	SNR $>1$ 97.4%, SNR $>3$ 91.0%
FLAME 3	$\|\Delta T\|$ ( ∘ C)	$29.89$	$42.04$	$59.58$	—
FLAME 3	$\|\Delta G\|$	$11.81$	$16.56$	$23.56$	—

Appendix

Table 21: physical-quantity measurements.

$\alpha$	true	random
0.00	$0.5005\pm 0.0433$	$0.5223\pm 0.0029$
0.25	$0.4879\pm 0.0290$	$0.5199\pm 0.0007$
0.50	$0.4827\pm 0.0260$	$0.5157\pm 0.0039$
0.75	$0.4776\pm 0.0231$	$0.5148\pm 0.0046$
1.00	$0.4720\pm 0.0207$	$0.5133\pm 0.0062$
1.25	$0.4665\pm 0.0187$	$0.5116\pm 0.0076$

Appendix

Table 22: [email protected] as a function of modulation amplitude

\alpha

over seeds s0, s1, and s2. All points except

\alpha=1.0

are out-of-distribution extrapolations.

Router	Set	$\nu_{T}^{K}$ / 152	per-image fraction (mean $\pm$ std)	max
fixed ( $\tau_{\text{high}}{=}0.7$ , $\tau_{\text{low}}{=}0.3$ )	train	132	$0.031\pm 0.072$	0.382
quantile ( $Q_{0.9}$ , $Q_{0.3}$ )	train	147	$0.092\pm 0.118$	0.428
fixed	test	138	$0.069\pm 0.110$	0.421
quantile	test	148	$0.152\pm 0.131$	0.467

Appendix

Table 23: Per-image switch fraction under the fixed-threshold and quantile routers, on the training and test evaluation sets.

\nu_{T}^{K}

is the aggregate count out of 152 checkpoint pairs. The per-image fraction appears alongside.

Infrared Object Detection

Infrared object detection locates and classifies objects in thermal imagery, including small, low-contrast targets and visible–thermal scenes.

See all 33 papers

Aug 11, 2026

Bridging Severe Cross-Modal Misalignment: End-to-End Visible-Infrared Object Detection via Explicit Feature-Domain Affine Registration

Visible-infrared object detection relies on complementary RGB and thermal cues, but its performance is often degraded by cross-modal spatial misalignment. Most existing methods rely on implicit feature adaptation to handle weakly misaligned scenarios, while large-offset geometric discrepancies remain insufficiently addressed. In this paper, we propose a Joint Feature-domain Registration and Detection network (JFRDet), an end-to-end visible-infrared oriented object detector tailored for severely cross-modal geometric discrepancies. JFRDet introduces a Cross-Modal Affine Alignment (CMAA) module to estimate an image-level affine transformation for explicit multi-level feature alignment. Note that illumination changes directly affect the reliability of RGB cues, an Illumination-Guided Complementary Fusion (IGCF) module adaptively exploits modality reliability under varying illumination conditions for cross-modal fusion. Then, an Alignment Quality-Consistency Gating (AQCG) strategy stabilizes joint optimization by modulating detection supervision according to alignment reliability and gradient consistency. We further construct DroneVehicle Misaligned (DVMA), a benchmark for evaluating visible-infrared oriented object detection under severe cross-modal geometric misalignment. The proposed JFRDet achieves 69.7%

\mathrm{mAP}_{50}

on DVMA, which represents state-of-the-art (SOTA) performance. The code and dataset will be available on GitHub.

Qi Ming, Yuyang Wang, Mingjing Zhao +7

Beijing University of Technology, China · Central South University, China · Beijing Electronics Science & Technology Institute +4

Multispectral Object Detection Cross-Modal Alignment Infrared Object Detection Cross-Modal Image Registration Multimodal Sensor Fusion

Ask PDF

Mar 23, 2026

Thermal Topology Collapse: Universal Physical Patch Attacks on Infrared Vision Systems

Infrared pedestrian detectors are increasingly deployed in all-weather perception systems, but their robustness against physical adversarial attacks remains insufficiently understood. Existing infrared physical attacks are mostly instance-specific, requiring perturbations to be optimized for particular samples, poses, or scenes, which limits scalability under changing deployment conditions. This paper proposes Universal Physical Patch Attack (UPPA), a universal cold-patch framework for infrared pedestrian detection. UPPA is built on the observation that infrared physical perturbations should exploit smooth, low-frequency thermal structures rather than visible-light texture patterns. It represents the attack carrier as topology-constrained Bézier Curved-Blocks, providing a compact and manufacturable geometric parameterization, and optimizes one shared perturbation with Particle Swarm Optimization (PSO) under Thin Plate Spline (TPS) deformation and Expectation over Transformation (EOT) imaging transformations. The optimized pattern is then deployed as wearable cold patches without sample-specific re-optimization during deployment. Experiments on five infrared datasets and nine pedestrian detectors show consistent digital attack performance, strong cross-dataset and cross-model transferability, and a 92.59% attack success rate in real-world physical experiments. Ablation, visualization, and defense analyses show that Curved-Blocks disrupt pedestrian thermal feature aggregation and remain difficult for image-restoration-style defenses to remove. These results reveal a practical universal physical vulnerability in current infrared pedestrian detection systems and provide a benchmark for robustness evaluation.

Chengyin Hu, Yikun Guo, Yuxian Dong +7

China University of Petroleum-Beijing at Karamay, No. 355, Anding Road, Karamay, 834000, Xinjiang, China. · Guangdong University of Technology, Guangzhou, Guangdong, China. · Tianjin University, Tianjin, China. +1

Adversarial Patch Attacks Physical Adversarial Attacks Adversarial Perturbations Adversarial Robustness Infrared Object Detection

Ask PDF

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

mIoU

P_d

, and

F_a

over state-of-the-art methods. We release our anonymous code and pretrained models.

Hao Li, Man Fung Zhuo

Department of Electrical and Computer Engineering, University of Arizona, Tucson, USA

Infrared Small Target Detection Infrared Object Detection Convolutional Neural Networks Semantic Segmentation

Ask PDF

Abstract

Infrared Object Detection

Explore similar work

Bridging Severe Cross-Modal Misalignment: End-to-End Visible-Infrared Object Detection via Explicit Feature-Domain Affine Registration

Thermal Topology Collapse: Universal Physical Patch Attacks on Infrared Vision Systems

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