Concentration, Not Uncertainty: Why Targeted Synthetic Data Doesn't Help Camouflaged Object Detection

Akshat Dobhal; Sanjay Singh

cs.CVOct 7, 2026

Concentration, Not Uncertainty: Why Targeted Synthetic Data Doesn't Help Camouflaged Object Detection

Authors: Akshat Dobhal, Sanjay Singh

Organizations: Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Abstract

Camouflaged object detection requires pixel-accurate masks, but obtaining such annotations is slow and costly, making synthetic training images an attractive alternative. Under a fixed generation budget, however, it remains unclear which real-image regions to target for synthetic data generation. We study an uncertainty-guided generation strategy that clusters the unlabelled real images, identifies clusters on which the model is least certain, allocates synthetic generation toward those clusters, and iteratively retrains the model. Across 103 training runs, uncertainty-based targeting does not outperform random allocation. Five independent controls further show that this null result is not an artifact: targeted training sets are measurably different from random sets, but the difference is explained by concentrating the generation budget rather than by where uncertainty is concentrated, as every concentration rule we test reproduces the effect and, on boundary accuracy, so does aiming at the clusters the model was most certain about. Separately, we find substantial data contamination in CHAMELEON, with 50 of its 76 images duplicated from training data despite the standard overlap check reporting zero overlap. Together, these results show that, under a fixed synthetic-data budget, budget concentration, not uncertainty-based targeting, accounts for the observed training-set effects.

Figures & tables

Arm	What it appends	Concentration	Direction
A0	nothing	—	—
A2	$1000$ real photographs	—	—
B	$1000$ renders drawn uniformly at random	dispersed	none
C10	the $1000$ the allocation selects	concentrated	as scored
CSHUF	C10’s quotas, cluster scores permuted	fixed, same shape	destroyed
CINV	C10’s quotas, their rank reversed	fixed, same shape	reversed

Table 1: Every arm shares one base pool and differs only in what is appended to it. The last two columns are the axes of Figure 1 : how concentrated the allocation is, and where it points. The C-family are permutations of a single allocation, so they hold concentration fixed and differ only in direction; the digits in C10 record the allocation temperature,

\tau=1.0

. None of A0, A2 and MT has an allocation to place.

Figure 1: What each arm holds fixed and what it varies. The horizontal axis is how concentrated an allocation is; the vertical axis is where it points, measured as rank correlation with the committed uncertainty ranking. Because the C-family are permutations of one allocation, they reproduce its shape to five decimal places and differ only vertically, which is what makes the vertical contrast clean and the diagonal one ambiguous. A0 and A2 append nothing and no generated content, respectively, so neither has an allocation to place. Sources: abc_pools.json ; T2_RESULTS.md ; results.md §5.2.

	What it tests	$n$	Gap	$\Delta$	$2\hat{\sigma}$	$\times$ bar	Verdict
ABC	targeting vs. random, as committed	$3$	C10 $-$ B	$+0.0050$	$0.0179$	$0.28$	within noise
SE	the same gap at eight seeds	$8$	C10 $-$ B	$+0.0024$	$0.0069$	$0.34$	within noise
OR	a perfect score in place of ours	$3$	CORACLE $-$ B	$+0.0008$	$0.0061$	$0.14$	within noise
FX	the optimisation budget unpinned	$3$	C10FX $-$ BFX	$+0.0014$	$0.0116$	$0.12$	within noise
PC	can this design see anything?	$3$	C10 $-$ MT	$+0.0197$	$0.0060$	$\mathbf{3.31}$	detected

Table 2: Trained comparison on the primary architecture and endpoint. A gap must exceed its own campaign’s bar and agree in sign across seeds to count. Four campaigns attack the decisive gap from different directions and none of them moves it; the positive control, which compares against a method we expected to be clearly worse, clears its bar more than threefold. Per-cell tables for both architectures and both endpoints are in Appendix C . Sources: abc_verdict.json and its or/ , fx/ , se/ variants; pc_verdict.json .

Arch.	Metric	$2\hat{\sigma}$	$\mathrm{C10}\!-\!\mathrm{B}$	$\mathrm{CINV}\!-\!\mathrm{B}$	$\mathrm{C10}\!-\!\mathrm{CSHUF}$	$\mathrm{C10}\!-\!\mathrm{CINV}$
SINet	Boundary IoU	$0.0101$	$+0.0138$ 3/3	$+0.0135$ 3/3	$+0.0038$ 2/3	$+0.0004$ 2/3
SINet	Boundary $F$	$0.0151$	$+0.0190$ 3/3	$+0.0173$ 3/3	$+0.0045$ 2/3	$+0.0017$ 2/3
SINet	IoU	$0.0109$	$+0.0112$ 3/3	$+0.0127$ 3/3	$+0.0019$ 2/3	$-0.0016$ 2/3
SINet-v2	Boundary IoU	$0.0062$	$+0.0010$ 2/3	$+0.0025$ 2/3	$-0.0002$ 2/3	$-0.0015$ 3/3
SINet-v2	Boundary $F$	$0.0065$	$+0.0026$ 2/3	$+0.0050$ 3/3	$-0.0002$ 2/3	$-0.0025$ 3/3
SINet-v2	IoU	$0.0100$	$-0.0003$ 2/3	$+0.0006$ 2/3	$-0.0013$ 2/3	$-0.0010$ 2/3

Table 3: Arm differences on boundary metrics, COD10K-test, same pooled-

\hat{\sigma}

rule. Post-hoc: not pre-registered, and no verdict is re-decided. On SINet the concentration gap

\mathrm{C10}-\mathrm{B}

clears its bar — and so does the anti -targeted

\mathrm{CINV}-\mathrm{B}

, by as much — while both direction contrasts stay within noise. SINet-v2 resolves none of it. Source: diag_subgroups.json .

Figure 2: (a) How far a selected set sits from a random set of the same size, in held-out Cohen’s

d

. Every rule that keeps the allocation’s concentration reproduces the separation; only destroying concentration collapses it. Bars are means across the four cells at a budget of

N=250

; whiskers show the range. (b) Trained

S_{\alpha}

at eight seeds for the three arms that hold concentration fixed and vary only direction; points are seeds, rules are arm means, the band is the

\pm 2\hat{\sigma}

bar. Sources: c1_attribution.csv ; out/se/abc_metrics.csv .

Panel of Figure 3 — Figure 3: Two tiers of the same finding. Top: six of the $41$ same-dimension re-encodes the CHAMELEON image above, its training-pool partner below. Bottom: six of the $9$ recovered by geometric adjudication, which the same-dimension detector reports unchecked because a rescale or a crop moves an image out of its dimension group; note the differing pixel dimensions in each pair’s labels.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Work	Task	Acquisition signal	Generator control	Closed-loop	Abstention / control
LAKE-RED [ 41 ]	COD synthesis	none (uniform)	background only	no	—
S2R-COD [ 24 ]	COD adaptation	none (uniform)	background only	pseudo-labels	—
GAUDA [ 11 ]	surgical seg.	epistemic, per class	image + mask	yes, online	none reported
DisCL [ 23 ]	long-tail cls.	hardness, per sample	guidance strength	curriculum	none reported
SynQuE [ 3 ]	dataset ranking	proxy score, per set	n/a (picks sets)	no	none reported
This work	COD adaptation	disagreement, per cluster	background only	pseudo-labels	same-shape shuffle, rank reversal, oracle

Appendix

Table 4: Where this work sits among systems that steer a synthetic-data budget. Acquisition signal is what decides where the budget goes; generator control is what the method can actually vary in the generated image; diversity control is what stops the budget piling onto near-duplicates; closed-loop means the downstream model’s state feeds back into generation; abstention/control is whether a signal-destroying or same-shape control is run, which is what separates an effect of aiming from an effect of concentrating. The last column is the one this paper argues is missing from the literature, not a claim of superiority; our own entry records a null.

	Cell	$n$	Gap	$\Delta$	$2\hat{\sigma}$	$\times$ bar	sign
ABC	SINet $\cdot$ COD10K	3	C10 $-$ B	$+0.0050$	$0.0179$	$0.28$	3/3
	SINet $\cdot$ NC4K			$+0.0021$	$0.0083$	$0.25$	3/3
	SINet-v2 $\cdot$ COD10K			$-0.0004$	$0.0123$	$0.03$	2/3
	SINet-v2 $\cdot$ NC4K			$+0.0009$	$0.0105$	$0.09$	2/3
SE	SINet $\cdot$ COD10K	8	C10 $-$ B	$+0.0024$	$0.0069$	$0.34$	6/8
	SINet $\cdot$ NC4K			$+0.0003$	$0.0059$	$0.06$	4/8

Appendix

Table 5: The decisive gap in all four architectures-endpoint cells, per campaign. Each campaign is read against its own pre-registered bar, never another’s. Source: abc_verdict.json and its se/ , or/ , fx/ variants.

Arch.	Endpoint	Gap	$\Delta$	95% CI	TOST $p$	smallest $\delta$
SINet	COD10K	$\mathrm{C10}-\mathrm{B}$	$+0.0024$	$[-0.0020,+0.0067]$	0.1009	$0.0059$
SINet	COD10K	$\mathrm{C10}-\mathrm{CSHUF}$	$+0.0003$	$[-0.0041,+0.0047]$	0.0189	$0.0039$
SINet	COD10K	$\mathrm{C10}-\mathrm{CINV}$	$-0.0021$	$[-0.0066,+0.0023]$	0.0905	$0.0057$
SINet	COD10K	$\mathrm{CSHUF}-\mathrm{CINV}$	$-0.0024$	$[-0.0055,+0.0007]$	0.0486	$0.0050$
SINet	NC4K	$\mathrm{C10}-\mathrm{B}$	$+0.0003$	$[-0.0033,+0.0040]$	0.0084	$0.0033$
SINet	NC4K	$\mathrm{C10}-\mathrm{CSHUF}$	$+0.0007$	$[-0.0029,+0.0042]$	0.0106	$0.0036$

Appendix

Table 6: Post-hoc interval estimates at eight seeds per arm ,

S_{\alpha}

on both endpoints. These decide nothing : verdicts remain the pre-registered

2\hat{\sigma}

rule of Section 2 . Intervals are Welch on the two arms actually compared, so they are not inflated by an arm outside the comparison. TOST (two one-sided tests; [ 19 ] ) is run at

\delta=0.005

; bold marks the comparisons that reject, and the final column gives the smallest

\delta

at which each comparison could have concluded equivalence. Source: out/se/abc_metrics.csv , via figs/make_tost_table.py using make_results.py ’s welch_tost .

	dinoL224			dinoL518			clipL224
$k$	sil.	bs.ARI	sd.ARI	sil.	bs.ARI	sd.ARI	sil.	bs.ARI	sd.ARI
5	0.0538	0.712	0.795	0.0566	0.624	0.732	0.0568	0.953	0.983
10	0.0868	0.676	0.709	0.0923	0.598	0.770	0.0550	0.692	0.805
15	0.1053	0.608	0.714	0.1152	0.573	0.592	0.0486	0.575	0.670
20	0.1180	0.629	0.693	0.1326	0.506	0.524	0.0486	0.635	0.646
30	0.1378	0.596	0.642	0.1527	0.539	0.617	0.0488	0.609	0.631

Appendix

Table 7: Full clustering sweep over the target set: silhouette (mean over 10

k

-means seeds), bootstrap ARI and seed ARI, for every

k

in the declared grid and all three embedding spaces. The principled

k

per space is bold . No

k

in any space reaches a silhouette conventionally regarded as well-separated structure, and clipL224 falls almost monotonically, so its selection criterion returns the grid edge rather than an interior maximum — recorded as a FAIL of the declared guard. Source: b1_k_sweep_*.csv ; block EXP B1 #3 @ results/REBUILD_LOG.txt:922 .

arch	signal	agg.	$\rho(\mathrm{MAE})$	$\rho(1-S_{\alpha})$	$\rho(1-\mathrm{IoU})$	order	seeds
SINet-v2	ES	whole	$+0.7070\pm 0.0404$	$+0.5628\pm 0.0359$	$+0.4591\pm 0.0315$	PASS	10/10
SINet-v2	ensemble A0	whole	$+0.6191\pm 0.0675$	$+0.5483\pm 0.0409$	$+0.4540\pm 0.0451$	PASS	9/10
SINet-v2	ensemble CSHUF	whole	$+0.5576\pm 0.0690$	$+0.4488\pm 0.0580$	$+0.3370\pm 0.0577$	PASS	10/10
SINet-v2	entropy	whole	$+0.7393\pm 0.0309$	$+0.5404\pm 0.0408$	$+0.4070\pm 0.0395$	PASS	10/10
SINet-v2	ES	bound	$-0.5485\pm 0.0745$	$-0.1146\pm 0.0675$	$+0.0537\pm 0.0558$	FAIL	0/10
SINet-v2	ensemble A0	bound	$-0.1960\pm 0.0524$	$+0.0425\pm 0.0410$	$+0.0933\pm 0.0396$	FAIL	0/10

Appendix

Table 8: Full signal-generality measurement, all 16 rows: four signal variants across three signal families, two architectures, two aggregations.

\rho

at the committed seed,

\pm

sd over 10

k

-means seeds. The whole-image block passes the pre-registered ordering

8/8

; the boundary block passes

2/8

with four rows going negative, an outcome outside the pre-registered interpretation space, reported as it fell. Source: rebuild/T2C/out/t2c_table.csv ; block EXP T2C @ results/REBUILD_LOG.txt:2550 .

		raw		$\|\,z_{\mathrm{obj}}$		$\|\,z_{\mathrm{unc}}$
arch	signal	MAE	$1-S_{\alpha}$	MAE	$1-S_{\alpha}$	MAE	$1-S_{\alpha}$
SINet	ES	$+0.6284$	$+0.3433$	$+0.5761$	$+0.4988$	$+0.5518$	$+0.3897$
SINet	entropy	$+0.5892$	$+0.4267$	$+0.5505$	$+0.5517$	$+0.5060$	$+0.4426$
SINet	ensemble A0	$+0.6028$	$+0.2698$	$+0.5411$	$+0.4301$	$+0.5603$	$+0.2583$
SINet	ensemble CSHUF	$+0.5212$	$+0.4006$	$+0.5002$	$+0.4857$	$+0.4857$	$+0.3937$
SINet-v2	ES	$+0.7070$	$+0.5628$	$+0.7062$	$+0.6272$	$+0.4457$	$+0.4744$

Appendix

Table 9: The area control ( EXP AC ). First-order partial Spearman on the same cluster partition and floors as the committed measurement.

z_{\mathrm{obj}}

= per-cluster mean object area on the endpoint side;

z_{\mathrm{unc}}

= per-cluster mean uncertain-region area on the target side. Ordering survives

7/8

and

6/8

against a floor of

6/8

declared before the run. Note that partialling object area raises

\rho(1-S_{\alpha})

in every row — the opposite of the confound’s prediction. Source: rebuild/AC/out/ac_table.csv ; block EXP AC @ results/REBUILD_LOG.txt:2612 .

arch	arm	seed	endpoint	$S_{\alpha}$	MAE
SINet	A0	42	COD10K	0.715086	0.074960
SINet	A0	43	COD10K	0.681706	0.077044
SINet	A0	45	COD10K	0.706058	0.077347
SINet	A2	42	COD10K	0.705507	0.075994
SINet	A2	43	COD10K	0.708334	0.078489
SINet	A2	45	COD10K	0.709196	0.080858

Appendix

Table 10: Per-run metrics at full recorded precision, COD10K-test endpoint, for every run of the two frozen campaigns and the mean-teacher positive control —

39

of the

103

training runs in this work; Table 11 gives the same runs on NC4K. Arms B and C10 appear once; they were re-scored, not re-trained, for the falsification campaign and reproduced their committed values exactly at six decimal places. Source: rebuild/ABC/out/abc_metrics.csv , out/t2/abc_metrics.csv and rebuild/PC/out/pc_metrics.csv ; blocks EXP ABC #3 and EXP T2 #3.

arch	arm	seed	endpoint	$S_{\alpha}$	MAE
SINet	A0	42	NC4K	0.765121	0.085752
SINet	A0	43	NC4K	0.751545	0.088874
SINet	A0	45	NC4K	0.757700	0.090228
SINet	A2	42	NC4K	0.760908	0.088385
SINet	A2	43	NC4K	0.762933	0.087466
SINet	A2	45	NC4K	0.765789	0.090892

Appendix

Table 11: Per-run metrics at full recorded precision, NC4K endpoint, for the same

39

runs as Table 10 . Source: as Table 10 .

Cluster	Budget	% of $N$	$n_{\text{target}}$	CAMO %	$n_{\text{test}}$	$e_{j}$
30	194	19.4	21	95.2	2	0.0707
34	141	14.1	81	85.2	5	0.0674
63	74	7.4	18	66.7	3	0.0609
13	67	6.7	57	17.5	24	0.0599
23	29	2.9	43	41.9	20	0.0515
73	26	2.6	39	46.2	16	0.0505

Appendix

Table 12: The ten most-funded clusters under Equation 2 (

N=1000

\tau=1.0

), and what the budget is actually aimed at. The target pool is

24.8\%

CAMO, but

51.5\%

of the generation budget lands on CAMO images — a

2.08\times

enrichment (

2.14\times

under the committed assignment). The two largest quotas take

33.5\%

of the budget, are

90.2\%

CAMO, and between them hold

7

of the

2026

primary-endpoint images. Source: rebuild/DIAG/out/diag_clusters.json .

Explore similar work

Training-Free Metrics for Synthetic Object Detection Data: A Proxy for Detector Performance
Jun 18, 2026Myeongseok Nam, Donghun Yeo, Seungwook KimSynthetic Image Detection Synthetic Training Data
GenGenAI, South Korea
From Unity Simulation to Diffusion-Based Augmentation: Quantifying Dataset Balance for Robust Object Detection
Sep 29, 2026Mohamed Benkedadra, Aissa Saoudi, Maxime Gloesener +2Object Detection Generative Adversarial Network
ILIA, Université de Mons Mons, Belgium · Université Polytechnique Hauts-de-France Valenciennes, France · DeepILIA, ILIA, Université de Mons Mons, Belgium +1
LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection
May 11, 2026Abhishek Moturu, Anna Goldenberg, Babak TaatiSynthetic Data Synthetic Training Data
Department of Computer Science University of Toronto The Hospital for Sick Children UHN KITE Research Institute T-CAIREM Vector Institute · Department of Computer Science Department of Laboratory Medicine and Pathobiology University of Toronto The Hospital for Sick Children T-CAIREM Vector Institute · Department of Computer Science Institute of Biomedical Engineering University of Toronto Rehabilitation Sciences Institute UHN KITE Research Institute Vector Institute

Jun 18, 2026, cs.CVJ/K move · Enter open · S save

Training-Free Metrics for Synthetic Object Detection Data: A Proxy for Detector Performance

Myeongseok Nam, Donghun Yeo, Seungwook Kim

GenGenAI, South Korea

Synthetic Image Detection Synthetic Training Data Detection Performance Object Detection Training-Free

Open paper AskPDF arXiv

Synthetic images are increasingly used to augment scarce real data for object detection. However, not all synthetic sets help equally, and the only way to know a set's value is to train a detector on it, which is slow and demands dense annotation. We ask whether a training-free metric can instead rank candidate synthetic training sets by their downstream utility. Existing image-set metrics such as FID, KID, and MMD compare two feature distributions with a single global statistic, which we show is mis-specified for detection-data selection in two ways: it is blind to per-image composition (object count, box scale, class mix), and even at fixed composition its global averaging washes out the appearance differences that separate high-mAP pools from low-mAP ones. We propose Conditional-Composition Domain Match (CCDM), which converts any feature-space distance into a composition-stratified comparison, matching candidate and target within metadata-defined strata without training a detector. On COCO and VisDrone-DET, the best CCDM variant ranks 19 candidate training sets in strong agreement with YOLOv8 mAP (Spearman \r{ho} = 0.97 and 0.96), outperforming FID, KID, and MMD. Furthermore, CCDM holds when reference metadata comes from detector pseudo-labels rather than ground-truth boxes.

Abstract

Explore similar work

Training-Free Metrics for Synthetic Object Detection Data: A Proxy for Detector Performance

From Unity Simulation to Diffusion-Based Augmentation: Quantifying Dataset Balance for Robust Object Detection

LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection