Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction

Chunming He; Dingming Zhang; Fengyang Xiao; Zhiwen Cao; Sina Farsiu

cs.CVOct 21, 2026

Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction

Authors: Chunming He, Dingming Zhang, Fengyang Xiao, Zhiwen Cao, Sina Farsiu

Organizations: Department of Biomedical Engineering, Duke University, Durham, NC 27708 USA · Adobe Inc., San Jose, CA, USA

Abstract

Curriculum learning orders samples by difficulty, but difficulty alone does not establish supervision reliability for dense prediction. Easy samples can expose shortcuts, hard samples can be informative, and noisy or ambiguous samples can mislead training. Building on our conference method CurriSeg, we present PaceShift, a stabilize-then-challenge paradigm combining supervision-adaptive curriculum selection with guarded spectral post-training. A learning-dynamics view motivates separating supervision reliability from temporal update value and applying spectral challenges after stabilization. The first phase estimates reliability from available supervision, augmentation consistency, model disagreement, and prediction certainty. Temporal statistics guide sample weighting, while pixel-uncertainty weights reduce the early influence of uncertain regions. The second phase ranks predictions from multiple spectral views using supervision agreement, robustness, reference consistency, and penalties. A pairwise preference objective measured against the frozen Phase-I reference, together with an anchor, refines the prediction. The framework supports full, weak, semi-supervised, and model-derived supervision. Experiments across context-entangled content segmentation (CECS), broader dense-prediction tasks, and foundation-model and cross-modal adaptations show accuracy gains. Targeted diagnostics indicate improved pooled calibration and corruption robustness in the evaluated settings. PaceShift preserves the deployed architecture and adds no inference computation. Code will be publicly released.

Figures & tables

Fig. 2: Curriculum paradigms for dense prediction. A naive curriculum orders samples easy-to-hard by a static ground-truth difficulty and can strengthen reliance on fragile texture cues. CurriSeg adds temporal statistics and a stabilize-then-challenge schedule, but its score is defined by dense-mask error and its single-radius challenge is unguarded. PaceShift keeps the same overall two-phase training skeleton and changes two core components: a four-cue supervision-adaptive score

\rho_{i}

and a guarded multi-radius spectral challenge during post-training, without changing the deployed model architecture.

Fig. 3: Framework of PaceShift. Top (schematic): performance rises in Phase I, is perturbed by the spectral challenge at $T_{c}$ , and settles higher in Phase II. Phase I (SA-RCS) : four reliability cues form a supervision-adaptive difficulty $d_{i}=1-\rho_{i}$ . After a warm-up on all samples, a percentile curriculum selects $\mathcal{S}_{t}$ , and TSSW and PUE reweight the loss at sample and pixel level. Phase II (SBFT-P) : a multi-radius challenge yields candidates that a guard scores into a preferred and dispreferred mask, refined against the reference $\theta_{0}$ . No inference parameters are added.

Supervision sources	Trust $\tau_{\mathrm{trust}}$	$Q_{\mathrm{sup}}$ source
Dense human mask (full; semi, labeled part)	0.7	GT mask
Sparse human cue (scribble, point, box)	0.4	weak-cue evidence
Model-derived supervision (EMA teacher, CAM-derived pseudo-mask)	0.3	teacher / pseudo-mask agree

TABLE I: Supervision-trust presets. In Eq. 2 ,

a^{\mathrm{sup}}{=}\tau_{\mathrm{trust}}

and

(a^{\mathrm{cons}},a^{\mathrm{dis}},a^{\mathrm{ent}})=(1-\tau_{\mathrm{trust}})\,\pi

with the fixed template

\pi=(0.4,0.4,0.2)

, and the guard in Eq. 11 reuses

\tau_{\mathrm{trust}}

with the fixed penalty scale

\varepsilon

. Point supervision averages cue satisfaction with frozen-proposal agreement without changing the preset.

Methods	Backbone	CHAMELEON				CAMO				COD10K				NC4K				$\Delta$ (%)
Methods	Backbone	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
MGL-R [ 58 ]	ResNet50	0.031	0.825	0.917	0.891	0.088	0.738	0.812	0.775	0.035	0.680	0.851	0.814	0.053	0.778	0.867	0.833	—
PreyNet [ 59 ]	ResNet50	0.027	0.844	0.948	0.895	0.077	0.763	0.854	0.790	0.034	0.715	0.894	0.813	0.047	0.798	0.887	0.838	—
FEDER [ 10 ]	ResNet50	0.028	0.850	0.944	0.892	0.070	0.775	0.870	0.802	0.032	0.715	0.892	0.810	0.046	0.808	0.900	0.842	—
FEDER+ [ 17 ]	ResNet50	0.026	0.858	0.952	0.898	0.068	0.790	0.881	0.807	0.030	0.736	0.910	0.818	0.043	0.825	0.912	0.850	2.46 $\uparrow$
FEDER++ (Ours)	ResNet50	0.025	0.863	0.957	0.898	0.066	0.797	0.889	0.810	0.029	0.742	0.917	0.820	0.042	0.830	0.916	0.851	3.59 $\uparrow$

TABLE II: Camouflaged object detection. “+” and “++” denote training the baseline under CurriSeg [ 17 ] and PaceShift.

\Delta

: mean relative improvement over the baseline across all metrics and datasets, with the sign reversed for lower-is-better metrics so that positive values indicate improvement. Bold : best within each backbone block. † : our DINO V2-L retraining of VSCode under the protocol of this table. ‡ : evaluated by us from the authors’ released model under the protocol of this table.

Methods	Training Time (h)	Relative Time/Epoch	GPU Memory (G)	$\Delta$ (%)
FEDER	9.62	1.00 $\times$	1.53	–
FEDER+	6.84	1.02 $\times$	1.62	2.46 $\uparrow$
FEDER++	7.95	1.18 $\times$	1.81	3.59 $\uparrow$
FSEL	11.54	1.00 $\times$	2.83	–
FSEL+	5.96	1.03 $\times$	2.92	2.22 $\uparrow$
FSEL++	6.93	1.20 $\times$	3.15	3.28 $\uparrow$

TABLE III: Training cost under full supervision (ResNet50, batch size

2

, single RTX 4090). Vanilla baselines follow official schedules (e.g., FEDER:

100

epochs), while CurriSeg and PaceShift use

70

epochs. Relative Time/Epoch is wall-clock time per epoch normalized by the corresponding vanilla baseline. Peak memory uses max_memory_allocated , with SBFT-P views processed sequentially and the frozen reference resident in Phase II.

\Delta

is the mean relative accuracy gain from Table II .

Methods	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	max. $p$ (vs. base)
FEDER	0.715 $\pm$ .004	0.892 $\pm$ .003	0.810 $\pm$ .002	—
FEDER+	0.736 $\pm$ .003	0.910 $\pm$ .003	0.818 $\pm$ .002	$<$ 0.001
FEDER++	0.742 $\pm$ .003	0.917 $\pm$ .002	0.820 $\pm$ .001	$<$ 0.001
FSEL	0.722 $\pm$ .004	0.891 $\pm$ .003	0.822 $\pm$ .002	—
FSEL+	0.742 $\pm$ .003	0.909 $\pm$ .002	0.831 $\pm$ .001	$<$ 0.001
FSEL++	0.746 $\pm$ .003	0.916 $\pm$ .002	0.833 $\pm$ .001	$<$ 0.001

TABLE IV: Statistical significance on COD10K (ResNet50), mean

\pm

std over five paired runs with shared seeds. Two-sided paired

t

-tests report the maximum

p

-value over

F_{\beta}

E_{\phi}

, and

S_{\alpha}

. Tests of

++

against

+

are reported in the text.

Fig. 4: Visual comparisons on COD, PIS, TOD, and CDD using the RUN baseline and its CurriSeg- and PaceShift-trained variants. PaceShift better recovers concealed structures and boundaries, with fewer false positives in texture-rich backgrounds.

Methods	CVC-ColonDB			ETIS			PIS
Methods	mDice $\uparrow$	mIoU $\uparrow$	$S_{\alpha}$ $\uparrow$	mDice $\uparrow$	mIoU $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
PolypPVT [ 20 ]	0.808	0.727	0.865	0.787	0.706	0.871	—
CoInNet [ 84 ]	0.797	0.729	0.875	0.759	0.690	0.859	—
LSSNet [ 85 ]	0.820	0.741	0.867	0.779	0.701	0.867	—
MEGANet [ 8 ]	0.793	0.714	0.854	0.739	0.665	0.836	—
MEGANet+ [ 17 ]	0.815	0.736	0.866	0.755	0.682	0.848	2.24 $\uparrow$

TABLE V: Polyp image segmentation (PVT V2 backbone).

Methods	GDD			GSD			TOD
Methods	mIoU $\uparrow$	$F_{\beta}^{\max}$ $\uparrow$	$M$ $\downarrow$	mIoU $\uparrow$	$F_{\beta}^{\max}$ $\uparrow$	$M$ $\downarrow$	$\Delta$ (%)
EBLNet [ 86 ]	0.870	0.922	0.064	0.817	0.878	0.059	—
IEBAF [ 87 ]	0.887	0.944	0.056	0.861	0.926	0.049	—
GhostingNet [ 88 ]	0.893	0.943	0.054	0.838	0.904	0.055	—
RFENet [ 89 ]	0.886	0.938	0.057	0.865	0.931	0.048	—
RFENet+ [ 17 ]	0.893	0.952	0.052	0.875	0.939	0.045	3.22 $\uparrow$

TABLE VI: Transparent object detection (PVT V2 backbone).

Methods	$S_{\alpha}$ $\uparrow$	$M$ $\downarrow$	$E_{\phi}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{m}$ $\uparrow$	$\Delta$ (%)
SINet V2 [ 7 ]	0.551	0.102	0.567	0.223	0.248	—
HitNet [ 90 ]	0.563	0.118	0.564	0.298	0.298	—
OAFormer [ 91 ]	0.541	0.121	0.535	0.216	0.239	—
CamoFormer [ 22 ]	0.589	0.100	0.588	0.330	0.329	—
CamoFormer+ [ 17 ]	0.599	0.091	0.603	0.345	0.346	4.59 $\uparrow$
CamoFormer++ (Ours)	0.607	0.088	0.617	0.359	0.361	7.70 $\uparrow$

TABLE VII: Concealed defect detection on CDS2K.

Paradigms	Model	Backbone	Base ( $S_{\alpha}$ )	$\Delta$ (%, +)	$\Delta$ (%, ++)
Convolutional detectors
Enc.-dec.	SINet-V2 [ 7 ]	Res2Net50	0.815	1.76 $\uparrow$	4.23 $\uparrow$
Graph reason.	MGL-R [ 58 ]	ResNet50	0.814	2.49 $\uparrow$	4.77 $\uparrow$
Bio-inspired	PreyNet [ 59 ]	ResNet50	0.813	2.54 $\uparrow$	5.10 $\uparrow$
Edge $+$ freq.	FEDER [ 10 ]	ResNet50	0.810	3.05 $\uparrow$	4.30 $\uparrow$
Attention and transformer detectors

TABLE VIII: Generality across architectures on COD10K. “+” and “++” denote CurriSeg [ 17 ] and PaceShift. Base is the baseline’s

S_{\alpha}

, and the two

\Delta

columns report the mean relative gain of each recipe over its baseline across four metrics (

M

F_{\beta}

E_{\phi}

S_{\alpha}

). For UGTR, the

+

cell reports the standard CurriSeg recipe in [ 17 ] .

Tasks	Datasets	Baselines	Metrics	Base	+	++	$\Delta$ (%)
Tier 1: low-contrast detection with ambiguous annotation
IRSTD	IRSTD-1k [ 72 ]	IRSAM [ 92 ]	IoU $\uparrow$	73.7	75.5	77.2	4.75 $\uparrow$
Shadow det.	SBU [ 73 ]	Spider [ 2 ]	BER $\downarrow$	0.040	0.038	0.037	7.50 $\uparrow$
Tier 2: low-contrast clinical dense prediction
Skin lesion	ISIC 2018 [ 74 ]	DconnNet [ 93 ]	DSC $\uparrow$	90.4	90.7	92.8	2.65 $\uparrow$
Reti. OCT	RETOUCH [ 75 ]	DconnNet [ 93 ]	DSC $\uparrow$	87.7	87.5	89.6	2.17 $\uparrow$

TABLE IX: Generalization to broader dense prediction families. One primary metric per family. “+” and “++” denote the CurriSeg and PaceShift recipes on the listed baseline, and

\Delta

is the relative gain of “++” over it. Only the training strategy changes, and splits follow each baseline.

Strategies	Method source	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
Vanilla (RUN) [ 12 ]	ICML25	0.030	0.747	0.903	0.827	—
Loss curriculum	—	0.031	0.742	0.898	0.824	1.23 $\downarrow$
Co-teaching [ 98 ]	NeurIPS18	0.030	0.752	0.905	0.829	0.28 $\uparrow$
Meta-Weight-Net [ 99 ]	NeurIPS19	0.029	0.751	0.908	0.830	1.20 $\uparrow$
InfoBatch [ 36 ]	ICLR24	0.030	0.748	0.905	0.828	0.12 $\uparrow$
SeTa [ 37 ]	CVPR25	0.030	0.753	0.906	0.829	0.34 $\uparrow$

TABLE X: Online training strategies under full dataset access (RUN, ResNet50, COD10K) at a fixed training horizon. “SA-RCS only” uses the complete Phase-I recipe and removes only Phase II.

Fig. 5: One-shot coreset selection on COD10K (RUN, ResNet50), showing absolute

F_{\beta}

differences (

\times 10^{-2}

) from random selection at each retention rate. We compare the SA-RCS diagnostic with Moderate [ 101 ] , CCS [ 102 ] , Dyn-Unc [ 34 ] , TDDS [ 35 ] , and label-free ELFS [ 103 ] . All differences vanish at

100\%

retention, where selection is inactive. Vanilla achieves

F_{\beta}

0.747

Methods	Method source	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
WS-SAM [ 104 ]	NeurIPS23	0.035	0.724	0.865	0.808	—
SEE [ 42 ]	TPAMI25	0.033	0.735	0.866	0.814	—
SEE+ [ 17 ]	ICML26	0.031	0.746	0.878	0.821	2.45 $\uparrow$
SEE++ (Ours)	—	0.029	0.761	0.892	0.827	5.06 $\uparrow$
SCALER [ 43 ]	ECCV26	0.029	0.758	0.897	0.827	—
SCALER+ [ 17 ]	ICML26	0.028	0.767	0.906	0.829	1.47 $\uparrow$

TABLE XI: Weak supervision on COD10K (supervision: point, backbone: PVT V2, foundation model: SAM3). “

+

” denotes CurriSeg adapted to this regime as described in the Supp., and “

++

” denotes PaceShift.

Methods	Method source	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
CoSOD [ 105 ]	WACV24	0.053	0.674	0.807	0.766	—
SEE [ 42 ]	TPAMI25	0.043	0.713	0.825	0.771	—
SEE+ [ 17 ]	ICML26	0.041	0.723	0.836	0.776	2.01 $\uparrow$
SEE++ (Ours)	—	0.038	0.740	0.848	0.783	4.94 $\uparrow$
SCALER [ 43 ]	ECCV26	0.034	0.733	0.849	0.790	—
SCALER+ [ 17 ]	ICML26	0.033	0.739	0.855	0.793	1.21 $\uparrow$

TABLE XII: Semi-supervision on COD10K (

1/16

labeled training data, PVT V2, SAM3). “

+

” denotes CurriSeg adapted to this regime as detailed in the Supp., and “

++

” denotes PaceShift.

Criteria	Flag P/R $\uparrow$	Ret. PL $S_{\alpha}$ $\uparrow$	AUROC $\uparrow$	AUPRC $\uparrow$
Confidence filter	0.589	0.801	0.682	0.641
Entropy only	0.631	0.807	0.708	0.669
Peer small-loss	0.643	0.810	0.723	0.687
Dyn-Unc [ 34 ]	0.702	0.817	0.774	0.735
Pseudo-mask–SAM3 agreement	0.735	0.822	0.806	0.774
Prediction–SAM3 agreement	0.681	0.814	0.759	0.719

TABLE XIII: Corrupted pseudo-label detection on COD10K under the

1/16

semi-supervised setting. We corrupt

20\%

of the EMA-teacher pseudo-masks, and the frozen SAM3 proposal is never corrupted. Each criterion flags its lowest-scoring

20\%

, matching the corruption rate, so flagging precision and recall coincide (Flag P/R, random

0.200

). “Ret. PL

S_{\alpha}

” is the mean quality of the retained pseudo-labels against held-out ground truth, used only for evaluation. AUROC and AUPRC are threshold-free (random

0.50

and about

0.20

). “Pseudo-mask–SAM3 agreement” scores every evaluated pseudo-mask, corrupted or not, directly by its IoU with the frozen proposal, without the segmenter.

Methods	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
SAM-Adapter [ 49 ]	0.025	0.800	0.918	0.883	—
SAM-Adapter+ [ 17 ]	0.023	0.822	0.935	0.888	3.29 $\uparrow$
SAM-Adapter++ (Ours)	0.022	0.827	0.940	0.892	4.70 $\uparrow$
SAM2-Adapter [ 107 ]	0.018	0.848	0.950	0.899	—
SAM2-Adapter+ [ 17 ]	0.016	0.869	0.962	0.903	3.82 $\uparrow$
SAM2-Adapter++ (Ours)	0.015	0.876	0.969	0.906	5.69 $\uparrow$

TABLE XIV: Foundation-model adapters on COD10K under full supervision. “+” and “++” denote CurriSeg and PaceShift recipes.

Methods	Method source	VOC		COCO	$\Delta$ (%)
Methods	Method source	Val.	Test	Val.	$\Delta$ (%)
DIAL [ 109 ]	ECCV24	74.5	74.9	44.4	—
WeCLIP [ 110 ]	CVPR24	76.4	77.2	47.1	—
WeCLIP++ (Ours)	—	77.3	77.8	48.1	1.36 $\uparrow$
ExCEL [ 111 ]	CVPR25	78.4	78.5	50.3	—
ExCEL++ (Ours)	—	79.2	79.2	50.9	1.04 $\uparrow$

TABLE XV: VLM-based weakly supervised semantic segmentation (mIoU %). “++” adds our PaceShift to each CLIP-based baseline.

Methods	$cS_{\alpha}$ $\uparrow$	$cF^{w}_{\beta}$ $\uparrow$	$cM$ $\downarrow$	$cF_{\beta}$ $\uparrow$	$cE_{\phi}$ $\uparrow$	cIoU $\uparrow$	$\Delta$ (%)
FC-CLIP [ 112 ]	0.080	0.076	0.872	0.090	0.191	0.072	—
OVCoser [ 113 ]	0.579	0.490	0.336	0.520	0.616	0.443	—
OVCoser++ (Ours)	0.593	0.526	0.312	0.556	0.651	0.466	5.78 $\uparrow$
COCUS [ 114 ]	0.668	0.615	0.265	0.631	0.697	0.568	—
COCUS++ (Ours)	0.683	0.647	0.252	0.662	0.725	0.584	4.02 $\uparrow$

TABLE XVI: VLM-based open-vocabulary camouflaged object segmentation on OVCamo. The

c\cdot

measures are the open-vocabulary COS metrics and cIoU is the class-aware IoU.

Methods	$M$ $\downarrow$	$F^{w}_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$\Delta$ (%)
SINetV2-Ref [ 7 , 54 ]	0.033	0.700	0.888	0.823	—
R2CNet [ 54 ]	0.036	0.669	0.879	0.805	—
R2CNet++ (Ours)	0.033	0.692	0.896	0.815	3.74 $\uparrow$
RefOnce [ 115 ]	0.019	0.819	0.937	0.890	—
RefOnce++ (Ours)	0.018	0.832	0.951	0.896	2.26 $\uparrow$

TABLE XVII: Reference-guided camouflaged object segmentation on R2C7K.

RCS	TSSW	PUE	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$
—	—	—	0.021	0.810	0.941	0.878
✓	—	—	0.020	0.817	0.946	0.879
—	✓	—	0.021	0.814	0.943	0.879
—	—	✓	0.021	0.813	0.942	0.878
✓	✓	—	0.020	0.821	0.949	0.880
✓	—	✓	0.020	0.819	0.947	0.879

TABLE XVIII: Full factorial over Phase-I without post-training. RCS: robust curriculum selection, the percentile selection of Eq. 3 , TSSW: temporal statistics-based sample weighting, PUE: pixel uncertainty weighting. The all-on configuration corresponds to SA-RCS.

Ranking criteria	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$
Random	0.021	0.814	0.944	0.878
Loss easy-first	0.021	0.812	0.943	0.877
Loss hard-first	0.020	0.816	0.945	0.878
Confidence easy-first	0.021	0.813	0.943	0.877
RCS ranking (CurriSeg)	0.020	0.820	0.949	0.880
Static $\rho_{i}$	0.020	0.818	0.947	0.879

TABLE XIX: Curriculum ordering under a matched budget on COD10K.

(a) Reliability cues	(b) Temporal weighting	(c) Pixel weighting
$\rho_{i}$ variants $M$ $\downarrow$ $F_{\beta}$ $\uparrow$ $E_{\phi}$ $\uparrow$ $S_{\alpha}$ $\uparrow$ Supervision only 0.035 0.730 0.848 0.788 w/o supervision 0.033 0.740 0.855 0.792 w/o consistency 0.032 0.742 0.857 0.793 w/o disagreement 0.034 0.735 0.851 0.789 w/o entropy 0.032 0.744 0.859 0.795 Full $\rho_{i}$ 0.031 0.748 0.863 0.797	$\omega_{i}$ variants $M$ $\downarrow$ $F_{\beta}$ $\uparrow$ $E_{\phi}$ $\uparrow$ $S_{\alpha}$ $\uparrow$ CurriSeg TSSW 0.018 0.832 0.959 0.887 Mono. incr. $\sigma$ 0.018 0.829 0.957 0.885 Mono. decr. $\sigma$ 0.018 0.831 0.958 0.886 w/o $\omega^{\mathrm{out}}$ 0.017 0.833 0.960 0.887 Unimodal mean 0.017 0.834 0.961 0.888 Param.-free full 0.017 0.836 0.963 0.889	PUE variants $M$ $\downarrow$ $F_{\beta}$ $\uparrow$ $E_{\phi}$ $\uparrow$ $S_{\alpha}$ $\uparrow$ No PUE 0.018 0.830 0.958 0.886 Hard mask 0.018 0.829 0.956 0.884 Soft only 0.018 0.830 0.957 0.885 $+$ decay only 0.017 0.832 0.959 0.886 $+$ floor only 0.017 0.834 0.961 0.888 $+$ floor $+$ decay 0.017 0.836 0.963 0.889

TABLE XX: Inside the supervision-adaptive curriculum. Each Phase-I variant is inserted into the otherwise unchanged full PaceShift pipeline, including Phase II, so every panel reports final performance. (a) Reliability cues under

1/16

semi-supervision (SCALER, SAM3). (b) Temporal weighting shape under full supervision. (c) Pixel uncertainty weighting under full supervision.

Variants	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$S_{\alpha}$ c $\uparrow$	ECE $\downarrow$	$\mathrm{GT\text{-}CAG}$ $\downarrow$
Five runs: mean (first row), std (second row)
SA-RCS only (no post)	0.019	0.823	0.951	0.880	0.788	0.015	0.048
	$\pm$ .0003	$\pm$ .002	$\pm$ .002	$\pm$ .001	$\pm$ .003	$\pm$ .001	$\pm$ .001
Single- $r$ SBFT (conf.)	0.018	0.827	0.955	0.883	0.802	0.014	0.044
	$\pm$ .0003	$\pm$ .002	$\pm$ .002	$\pm$ .001	$\pm$ .003	$\pm$ .001	$\pm$ .001
Multi- $r$ supervised SBFT	0.018	0.829	0.957	0.884	0.811	0.013	0.040

TABLE XXI: Phase-II designs on COD10K (RUN, PVT V2). Upper part: mean over the five runs of Fig. 6 , with the sample standard deviation in the row beneath. Within each run, all post-training variants share the Phase-I endpoint and the optimization-step budget, and SA-RCS only evaluates that endpoint without post-training. Lower part: single-run controls from the canonical run (run 1). ECE is pixel-pooled,

\mathrm{GT\text{-}CAG}

is a confidence diagnostic, and

S_{\alpha}

c is the corruption-averaged

S_{\alpha}

. Guard drop-one ablations are in the Supp.

Training order	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$S_{\alpha}$ c $\uparrow$	$\mathrm{GT\text{-}CAG}$ $\downarrow$
Standard training	0.021	0.810	0.941	0.878	0.778	0.056
SBFT after warm-up	0.023	0.785	0.924	0.858	0.744	0.062
SBFT-P after warm-up	0.022	0.792	0.930	0.865	0.752	0.059
SBFT-P $\rightarrow$ SA-RCS	0.020	0.812	0.943	0.876	0.771	0.052
SA-RCS $\rightarrow$ SBFT (conf.)	0.018	0.828	0.956	0.884	0.803	0.043
SA-RCS $\rightarrow$ SBFT-P (ours)	0.017	0.836	0.963	0.889	0.832	0.035

TABLE XXII: Phase order on COD10K (single run). All schedules share the same warm-up, and for challenge-first variants the frozen warm-up checkpoint serves as the reference.

Trust design	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\phi}$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$S_{\alpha}$ (point) $\uparrow$
Uniform cue	0.034	0.731	0.848	0.786	0.820
Single $\tau_{\mathrm{trust}}{=}0.5$	0.033	0.738	0.854	0.791	0.826
Mismatched trust	0.036	0.719	0.838	0.777	0.812
Preset $\pm 20\%$	0.032 $\pm$ 0.001	0.745 $\pm$ 0.002	0.861 $\pm$ 0.002	0.796 $\pm$ 0.001	0.831 $\pm$ 0.001
Source-type preset	0.031	0.748	0.863	0.797	0.834
Val.-tuned $\tau_{\mathrm{trust}}$	0.031	0.749	0.864	0.798	0.835

TABLE XXIII: Trust parameterization (

1/16

semi-supervision, plus point supervision for the sparse-human preset). Uniform: equal cue coefficients, single

\tau_{\mathrm{trust}}

: one shared trust for all sources, “Mismatched”: supervision receives an incorrect source-type preset (in semi the dense-human and model-derived presets are swapped, and in point the sparse-human source receives the model-derived preset). The

\pm 20\%

row reports mean

\pm

std over ten order-preserving random draws. Variation is across configurations, not seeds. Per-regime tuning selects

\tau_{\mathrm{trust}}

on held-out training data only.

Fig. 6: Training dynamics on COD10K (RUN with PVT V2):

F_{\beta}

on the test set at each checkpoint, mean over five runs with one-standard-deviation bands. The marks on the epoch axis indicate

T_{w}

T_{c}

, and

T

, and ten extra diagnostic epochs beyond

T

are shown for reference.

Fig. 7: Robustness on COD10K as retention of each recipe’s own clean

S_{\alpha}

. Severity

0

is the clean test set. At each severity, retention is averaged within each corruption family and then equally over the four families.

(a) Recipe-level comparison on COD10K
Metrics	RUN	$+$ TS	$+$ CurriSeg	$+$ PaceShift	Direct conf. reward
ECE $\downarrow$	0.019	0.010	0.016	0.011	0.024
$\mathrm{GT\text{-}CAG}$ $\downarrow$	0.056	0.034	0.046	0.035	0.071
$\mathrm{CAG}_{\mathrm{bnd}}$ $\downarrow$	0.121	0.102	0.096	0.074	0.148
WConf $\downarrow$	0.86	0.80	0.82	0.79	0.94

TABLE XXIV: Calibration and confidence diagnostics. ECE: pixel-pooled expected calibration error, used to assess calibration.

\mathrm{GT\text{-}CAG}

: ground-truth-stratified confidence–accuracy gap, the mean of the gaps on ground-truth foreground and background pixels,

\mathrm{CAG}_{\mathrm{bnd}}

: the same gap within a dilated boundary band, and WConf: mean confidence on incorrect pixels, all three being complementary diagnostics. TS [ 83 ] : post-hoc temperature scaling, a single scalar fitted on held-out validation data. (a) Recipe-level comparison on COD10K (single run), where the final column is the direct confidence-reward control defined in the Supp. (b) The guarded stage across tasks and supervision regimes (RUN with PVT V2 for PIS and TOD, SCALER for the COD regimes).

(a) $\rho_{i}$ vs. pseudo-label quality	(b) Alignment with $A^{\mathrm{opt}}$
$\rho_{i}$ quin. IoU $\uparrow$ bnd. F $\uparrow$ pix. err. $\downarrow$ Q5 (high) 0.871 0.822 0.032 Q4 0.804 0.741 0.058 Q3 0.742 0.668 0.083 Q2 0.668 0.585 0.121 Q1 (low) 0.549 0.468 0.192	Scores Spearman $\uparrow$ Top- $20\%$ $\uparrow$ Instant. loss $-$ 0.25 $\pm$ 0.02 11 $\pm$ 1 % GraNd 0.20 $\pm$ 0.01 23 $\pm$ 1 % Forgetting 0.29 $\pm$ 0.01 25 $\pm$ 1 % $\rho_{i}$ 0.32 $\pm$ 0.01 28 $\pm$ 1 % $\omega_{i}$ 0.53 $\pm$ 0.02 41 $\pm$ 2 %

TABLE XXV: Validation of PaceShift’s internal training signals. (a) Pseudo-label quality across

\rho_{i}

quintiles on the unlabeled split of the semi-supervised

1/16

regime. Ground truth is withheld from training and used only post hoc, and

r_{s}{=}0.64

is computed over individual pseudo-labels before binning. (b) Alignment of sample scores on

\mathcal{D}_{s}

with the update-direction proxy

A^{\mathrm{opt}}

: Spearman and top-

20\%

overlap (random:

20\%

), reported as mean

\pm

std over the four Phase-I refresh checkpoints with at least two history entries. Variation is across checkpoints, not training seeds.

Selection rules	Semi $1/16$ (unlab.)		Point
Selection rules	Top-1 (%) $\uparrow$	Regret $\downarrow$	Top-1 (%) $\uparrow$	Regret $\downarrow$
Random (expected)	36.7	0.052	34.9	0.061
Confidence only	47.1	0.038	44.8	0.046
Consistency only	54.6	0.031	52.9	0.038
Guard w/o $Q_{\mathrm{sup}}$	58.3	0.027	57.1	0.033
Full guard $Q(m)$	71.8	0.014	68.4	0.019

TABLE XXVI: Guard selection against held-out GT, shown only where

Q_{\mathrm{sup}}

is surrogate-based. Top-1 is the rate of selecting a candidate within

10^{-3}

true IoU of the oracle candidate, and Regret is the true IoU left on the table. Random reports the exact expectation under uniform selection among the three candidates, including tolerance-induced oracle ties.

Training	Clean $\uparrow$	drop under intervention
Training	Clean $\uparrow$	HF rem. $\downarrow$	HF noise $\downarrow$	LF supp. (DC retained)
Standard	0.878	0.086	0.177	0.276
Naive easy-first CL	0.871	0.104	0.201	0.249
SA-RCS only	0.881	0.071	0.142	0.288
PaceShift	0.889	0.049	0.098	0.309

TABLE XXVII: Test-time frequency intervention on COD10K (RUN, PVT V2). Each cell is the drop from that model’s own clean

S_{\alpha}

. The naive easy-first row is a conventional easy-to-hard schedule, unlike the matched-budget control of Table XIX .

Curriculum Learning

Curriculum learning organizes training examples, tasks, or environments to match how a model learns.

See all 172 papers

Jun 26, 2026

Curriculum-guided Change Detection Training: Toward Accurate Serac Fall Monitoring

Change Detection (CD) aims to identify semantic or structural changes from nearly registered multi-temporal images. While recent advances in training methodologies have largely focused on semi-supervised learning and consistency regularization, alternative training paradigms remain underexplored. In particular, most deep CD methods rely on uniform sampling during training, implicitly assuming that all training samples contribute equally to the optimization process. However, such naive sampling can introduce noisy gradients and hinder robust representation learning. To address this limitation, we propose a curriculum learning framework tailored for change detection. Our approach investigates two complementary difficulty measures: the Solar Angular Gap (SAG), a physically grounded proxy for acquisition-condition variability, and the Structural Similarity Index Measure (SSIM), which evaluates appearance similarity between image pairs. Based on these criteria, the framework progressively introduces challenging samples during training, enabling models to learn robust representations in a coarse-to-fine manner. We evaluate our method on the challenging SeracFallDet benchmark, where results demonstrate consistent improvements of the proposed approach over standard uniform-sampling strategies for both pixel-based and object-based approaches. These results highlight the potential of curriculum learning to improve robustness in deep change detection. Importantly, our training framework is orthogonal to existing CD architectures, making it readily applicable to a broad range of methods.

Arthur Dérédel, Carlos Crispim-Junior, Pierre Lemaire +2

LIRIS, UMR5205, Universit´e Lumi`ere Lyon 2, CNRS, Ecole Centrale de Lyon, INSA Lyon, Universit´e Claude Bernard Lyon 1, 69676, Bron, France · Styx4D, 19 rue lac Saint Andr´e, Le Bourget-du-Lac, 73370, France

Remote Sensing Change Detection Curriculum Learning

Ask PDF

Aug 29, 2026

GramLoop: Training-Free Gram-Gated Replay for Robust Dense Prediction

We aim to improve frozen DINOv3 dense-prediction models under distribution shift by adding inference computation inside the visual backbone, without changing model weights, task adapters, or prediction heads. The challenge is that repeated transformer-block computation must refine dense features without disrupting the pairwise patch relations that DINOv3 uses to preserve spatial structure. We introduce GramLoop, a training-free framework that replays a short transformer window and controls each replay through final-layer cosine-Gram consistency. Each proposal is propagated through the frozen suffix, measured against the standard DINOv3 trajectory, and accepted through a patchwise gate at the replay-window endpoint. Across object detection and semantic segmentation under corruptions, perturbations, and natural shifts, GramLoop improves all five shifted benchmarks over the paired DINOv3 baseline. On COCO-O, it improves mAP by +0.252 and Effective Robustness by +0.250, while preserving clean ADE20K performance. Code will be released at https://github.com/cheyan9/GramLoop.

Yang Chen, Canyu Shen, Xinzhe Rao +5

Huazhong University of Science and Technology · Tongji University · King Abdullah University of Science and Technology +2

Distribution Shift Robustness Test-Time Adaptation Vision Foundation Model Adaptation Object Detection Dense Prediction Semantic Segmentation

Ask PDF

May 19, 2026

Continual Segmentation under Joint Nonstationarity

Evolving data streams induce joint nonstationarity in continual semantic segmentation, where semantic classes, input distributions, and supervision availability change simultaneously over time. This setting reflects practical structured prediction systems, yet remains largely unexplored in prior continual learning work, which typically studies these factors in isolation. We formalize continual segmentation under coupled class, domain, and label shifts and investigate learning in heterogeneous dense prediction environments with limited annotations and abundant unlabeled data. To address instability and overfitting arising from few-shot supervision under distribution drift, we introduce gradient-adaptive stabilization, a parameter-wise regularization mechanism implemented via gradient-scaled stochastic perturbations that promotes a principled stability-plasticity tradeoff. We further leverage unlabeled data through semi-supervised learning and introduce prototype anchored supervision that validates pseudo-labels via joint confidence and prototype consistency. Together, these mechanisms enable learning under joint nonstationarity in continual segmentation. Extensive empirical evaluation across class-incremental, domain-incremental, and few-shot regimes demonstrates consistent improvements over prior methods in heterogeneous structured prediction settings. Our results expose fundamental failure modes of existing continual segmentation approaches and provide insight into learning robust dense predictors in dynamically evolving environments.

Prashant Pandey, Himanshu Kumar, Devineni Sri Venkatraya Chowdary +1

Bharti School of Telecommunications Technology & Management, Indian Institute of Technology, Delhi, India · Yardi School of Artificial Intelligence, Indian Institute of Technology, Delhi, India · Department of Computer Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India

Class-Incremental Learning Continual Learning Few-Shot Semantic Segmentation Semantic Segmentation

Ask PDF

Abstract

Curriculum Learning

Explore similar work

Curriculum-guided Change Detection Training: Toward Accurate Serac Fall Monitoring

GramLoop: Training-Free Gram-Gated Replay for Robust Dense Prediction

Continual Segmentation under Joint Nonstationarity