MC-PanDA++: Simpler, Stronger, and More Robust Domain-Adaptive Panoptic Segmentation

Ivan Martinović; Josip Šarić; Yuki M. Asano; Siniša Šegvić

cs.CVSep 30, 2026

MC-PanDA++: Simpler, Stronger, and More Robust Domain-Adaptive Panoptic Segmentation

Authors: Ivan Martinović, Josip Šarić, Yuki M. Asano, Siniša Šegvić

Organizations: Faculty of Electrical Engineering and Computing, University of Zagreb, Unska 3, 10000 Zagreb, Croatia. · Faculty of Computer and Information Science, University of Ljubljana, Večna Pot 113, 1000 Ljubljana, Slovenia. · Fundamental AI Lab, University of Technology Nuremberg, Dr.-Luise-Herzberg-Straße, 90461 Nürnberg, Germany.

Abstract

Unsupervised domain adaptation (UDA) reduces the annotation burden in panoptic segmentation by leveraging a cost-effectively labeled source domain (e.g., synthetic) and an unlabeled target domain to bridge the distribution gap. Existing panoptic UDA methods rely on teacher-student consistency learning built upon suboptimal per-pixel segmentation architectures. In contrast, state-of-the-art mask transformers are rarely adopted due to their pronounced vulnerability to confirmation bias in consistency learning, where erroneous teacher predictions are reinforced during training. Our earlier approach, MC-PanDA, mitigates this issue through fine-grained confidence estimation, which suppresses gradients from unreliable masks while sampling informative yet reliable locations for loss computation. However, this method entails a complex multi-stage training and requires careful hyperparameter tuning. This work presents MC-PanDA++, which addresses these limitations by introducing: (i) self-supervised vision encoders that provide a stronger and more robust initialization, further reducing the reliance on human annotations, (ii) per-class, self-adapting mask-wide loss scaling that stabilizes training and enables the usage of a single set of hyperparameters across domains, and (iii) a single-stage training pipeline that decreases overall conceptual complexity. Together, these improvements result in a conceptually simpler, better-performing, and more robust method for domain-adaptive panoptics. Source code: https://github.com/martinovicivan/MC-PanDA

Figures & tables

Figure 1: Validation Panoptic Quality (PQ) during domain adaptation from Cityscapes to ACDC. Baseline consistency training collapses due to confirmation bias, producing hallucinated false-positive masks, whereas our MC-PanDA ++ mitigates this issue through curated consistency learning based on mask confidence estimation.

Figure 2: Our UDA baseline jointly optimizes the student with a supervised source-domain loss and a target-domain teacher–student consistency loss, thereby exposing it to confirmation bias. MC-PanDA ++ mitigates this issue by estimating fine-grained teacher confidence, which guides mask-wide loss scaling and point sampling for loss computation.

Figure 3: Panoptic performance (PQ) on Synthia

\rightarrow

Vistas as a function of the initial threshold

\tau_{1}

. Each point denotes the mean PQ over three random seeds, with error bars indicating the standard deviation. The orange point marks the default MC-PanDA setting with

\tau_{1}=0.99

Figure 4: MC-PanDA ++ simplifies the MC-PanDA training pipeline by removing source-domain pre-training and the teacher burn-in stage.

	Synthia $\rightarrow$ City			Synthia $\rightarrow$ Vistas
Method	$\textbf{PQ}_{16}$	$\text{RQ}_{16}$	$\text{SQ}_{16}$	$\textbf{PQ}_{16}$	$\text{RQ}_{16}$	$\text{SQ}_{16}$
CVRN ( Huang et al., 2021 )	32.1	40.9	66.6	21.3	28.1	65.3
UniDAF-DETR ( Zhang et al., 2023 )	33.0	42.2	64.7	–	–	–
UniDAF-PSN ( Zhang et al., 2023 ; Kirillov et al., 2019b )	34.2	44.3	66.9	–	–	–
EDAPS ( Saha et al., 2023 )	41.2	53.6	72.7	36.6	46.1	71.7
EDAPS † ( Saha et al., 2023 )	39.3	51.4	73.1	–	–	–

Table 1: Comparison with the state of the art on Synthia

\rightarrow

Cityscapes and Synthia

\rightarrow

Vistas.

\dagger

indicates our reimplementation of EDAPS using the same Swin-B backbone as in MC-PanDA.

\ddagger

indicates our reimplementation with the same DINOv2-B backbone as in MC-PanDA ++ . All experiments for MC-PanDA and MC-PanDA ++ are averaged over three random seeds.

	Cityscapes $\rightarrow$ Foggy				Cityscapes $\rightarrow$ Vistas
Method	$\textbf{PQ}_{16}$	$\text{RQ}_{16}$	$\text{SQ}_{16}$	$\textbf{PQ}_{19}$	$\textbf{PQ}_{16}$	$\text{RQ}_{16}$	$\text{SQ}_{16}$	$\textbf{PQ}_{19}$
CVRN ( Huang et al., 2021 )	35.7	46.7	72.7	–	33.5	42.8	73.8	–
UniDAF ( Zhang et al., 2023 )	37.6	49.5	72.9	–	–	–	–	–
EDAPS ( Saha et al., 2023 )	56.7	70.5	79.2	–	41.2	53.4	75.9	–
LIDAPS ( Mansour et al., 2025 )	59.6	73.2	80.2		42.6	54.9	76.6
MC-PanDA ( Martinović et al., 2024 )	63.7	76.3	82.5	62.0	53.8	66.6	79.3	51.7

Table 2: Performance evaluation on Cityscapes

\rightarrow

Foggy Cityscapes and Cityscapes

\rightarrow

Vistas. All experiments for MC-PanDA and MC-PanDA ++ are averaged over three random seeds.

			SYN $\rightarrow$ City		SYN $\rightarrow$ Vistas
BC	$\text{MLS}_{.99}$	CBPF	$\text{PQ}_{16}$		$\text{PQ}_{16}$
–	–	–	30.8	$\Rsh$	25.2	$\Rsh$
✓	–	–	$\text{39.6}_{\pm 1.5}$	+8.8	$\text{32.2}_{\pm 0.9}$	+7.0
✓	✓	–	$\text{44.6}_{\pm 0.7}$	+13.8	$\text{37.1}_{\pm 1.0}$	+11.9
✓	–	✓	$\text{44.1}_{\pm 0.2}$	+13.3	$\text{35.1}_{\pm 1.3}$	+9.9
✓	✓	✓	$\text{47.4}_{\pm 0.8}$	+16.6	$\text{38.7}_{\pm 1.0}$	+13.5

Table 3: Ablation study on Synthia

\rightarrow

Cityscapes and Synthia

\rightarrow

Vistas with the Swin-B backbone, as originally reported for MC-PanDA ( Martinović et al., 2024 ) . Top row corresponds to supervised training on Synthia. We report mean

{}_{\pm\text{std}}

over three random seeds. BC: baseline consistency (section 3.2 ).

ILS	MLS	CBPF	SQ 16	RQ 16	PQ 16
✓	–	–	$\text{73.7}_{\pm 0.5}$	$\text{50.5}_{\pm 1.1}$	$\text{39.7}_{\pm 0.8}$
–	✓	–	$\text{75.3}_{\pm 0.2}$	$\text{56.5}_{\pm 1.0}$	$\text{44.7}_{\pm 0.7}$
✓	–	✓	$\text{74.8}_{\pm 0.4}$	$\text{53.4}_{\pm 0.2}$	$\text{42.1}_{\pm 0.5}$
–	✓	✓	$\text{76.7}_{\pm 0.4}$	$\text{59.3}_{\pm 1.0}$	$\text{47.4}_{\pm 0.8}$

Table 4: Comparison of image-wide (ILS) and mask-wide loss scaling (MLS) on Synthia

\rightarrow

Cityscapes, as originally reported for MC-PanDA ( Martinović et al., 2024 ) . We report an optimistic ILS performance as the maximum across three different

\tau_{\text{ILS}}

thresholds.

CBPF ( $\tau_{2}=0.8)$
filtering method	$\text{SQ}_{16}$	$\text{RQ}_{16}$	$\textbf{PQ}_{16}$
random sampling	$\text{74.9}_{\pm 0.1}$	$\text{56.1}_{\pm 1.4}$	$\text{44.2}_{\pm 1.1}$
per-mask	$\text{75.2}_{\pm 0.4}$	$\text{56.5}_{\pm 0.1}$	$\text{44.7}_{\pm 0.1}$
$\text{all-masks}~(\text{eq.}\ \ref{eq:conf_def})$	$\text{76.7}_{\pm 0.4}$	$\text{59.3}_{\pm 1.0}$	$\text{47.4}_{\pm 0.8}$

Table 5: Validation of loss subsampling with confidence-based point filtering (CBPF) on Synthia

\rightarrow

Cityscapes, as originally reported for MC-PanDA ( Martinović et al., 2024 ) . The random sampling baseline applies the loss to a random subset of image points.

Figure 5: Comparison of initialization strategies. MC-PanDA with a self-supervised ViT-B/16 (DINO, IN-1k) yields the best PQ, outperforming supervised ViT-B/16 models pre-trained on ImageNet-1k and even ImageNet-21k. Results are averaged over three seeds.

Figure 6: Selection of the pretrained backbone. We compare different pretraining strategies in a domain-generalization setting across four benchmarks: reconstruction-based self-supervised pretraining (MAE), vision-language pretraining (SigLIP2), and distillation-based self-supervised pretraining (Franca, DINOv2, and DINOv3).

			Syn $\rightarrow$ City		Syn $\rightarrow$ Vistas
BC	$\text{MLS}_{.99}$	$\text{CBPF}_{0.8}$	$\textbf{PQ}_{16}$		$\textbf{PQ}_{16}$
–	–	–	36.0	$\Rsh$	33.7	$\Rsh$
✓	–	–	$\text{39.7}_{\pm 0.9}$	+3.7	$\text{33.6}_{\pm 1.6}$	-0.1
✓	✓	–	$\text{44.5}_{\pm 0.6}$	+8.5	$\text{42.8}_{\pm 0.4}$	+9.1
✓	✓	✓	$\text{48.5}_{\pm 0.4}$	+12.5	$\text{44.2}_{\pm 0.1}$	+10.5

Table 6: Ablation study on Synthia

\rightarrow

Cityscapes and Synthia

\rightarrow

Vistas with the self-supervised DINOv2 initialization. Top row corresponds to supervised training on Synthia. We report mean

{}_{\pm\text{std}}

over three random seeds.

Panel of Figure 7 — Figure 7: Visual comparison between the synthetic datasets Synthia ( Ros et al., 2016 ) (top row) and UrbanSyn ( Gómez et al., 2025 ) (bottom row).

			USyn $\rightarrow$ City		USyn $\rightarrow$ Vistas
BC	$\text{MLS}_{.99}$	$\text{CBPF}_{0.8}$	$\textbf{PQ}_{19}$		$\textbf{PQ}_{19}$
–	–	–	48.3	$\Rsh$	42.1	$\Rsh$
✓	–	–	$\text{53.4}_{\pm 0.8}$	+5.1	$\text{45.3}_{\pm 1.2}$	+3.2
✓	✓	–	$\text{54.5}_{\pm 0.4}$	+6.2	$\text{47.3}_{\pm 0.1}$	+5.2
✓	✓	✓	$\text{56.8}_{\pm 0.7}$	+8.5	$\text{48.7}_{\pm 0.6}$	+6.6

Table 7: Ablation study on UrbanSyn

\rightarrow

Cityscapes and UrbanSyn

\rightarrow

Vistas with the self-supervised DINOv2 initialization. Top row corresponds to supervised training on UrbanSyn. We report mean

{}_{\pm\text{std}}

over three random seeds. USyn: UrbanSyn.

Src. domain	$\text{PQ}_{16}$ on City		$\text{PQ}_{16}$ on Vistas
Synthia	$\text{48.5}_{\pm 0.4}$	$\Rsh$	$\text{44.2}_{\pm 0.1}$	$\Rsh$
UrbanSyn	$\text{59.8}_{\pm 0.4}$	+11.3	$\text{52.2}_{\pm 0.3}$	+8.0

Table 8: Comparison of MC-PanDA with DINOv2 when using Synthia versus UrbanSyn as the synthetic source domain. Results are reported in

\text{PQ}_{16}

for the 16 classes shared between both datasets.

Panel of Figure 8 — Figure 8: Example images from the adverse-condition datasets: ACDC ( Sakaridis et al., 2021 ) (top row) and MUSES ( Brödermann et al., 2024 ) (bottom row).

			City $\rightarrow$ ACDC		City $\rightarrow$ MUSES
BC	$\text{MLS}_{.99}$	$\text{CBPF}_{0.8}$	$\textbf{PQ}_{19}$		$\textbf{PQ}_{19}$
–	–	–	48.7	$\Rsh$	42.0	$\Rsh$
✓	–	–	$\text{43.3}_{\pm 1.2}$	-5.5	$\text{49.3}_{\pm 0.5}$	+7.3
✓	✓	–	$\text{53.5}_{\pm 1.1}$	+4.8	$\text{50.6}_{\pm 0.7}$	+8.7
✓	✓	✓	$\text{55.5}_{\pm 1.0}$	+6.8	$\text{52.6}_{\pm 0.2}$	+10.6

Table 9: Ablation study on Cityscapes

\rightarrow

ACDC and Cityscapes

\rightarrow

MUSES the self-supervised DINOv2 initialization. Top row corresponds to supervised training on Cityscapes. We report mean

{}_{\pm\text{std}}

over three random seeds.

# stages	Synthia $\rightarrow$ Vistas		City $\rightarrow$ ACDC
# stages	$\tau_{1}=0.9$	$\tau_{1}=0.99$	$\tau_{1}=0.9$	$\tau_{1}=0.99$
three	$\text{42.1}_{\pm 3.3}$	$\text{44.2}_{\pm 0.1}$	$\text{52.9}_{\pm 1.2}$	$\text{55.5}_{\pm 1.0}$
single	$\text{42.3}_{\pm 1.7}$	$\text{44.4}_{\pm 1.2}$	$\text{54.4}_{\pm 1.1}$	$\text{54.1}_{\pm 1.5}$

Table 10: Comparison of the original three-stage MC-PanDA pipeline and the simplified single-stage MC-PanDA ++ pipeline using a constant

\tau_{1}

Checkpoint	MC-PanDA three-stage	MC-PanDA single-stage
50k	40.8	41.5
70k	42.8	43.0
90k	43.7	43.9
110k	44.2	44.4

Table N1: Convergence comparison on Synthia

\rightarrow

Vistas. Both variants use DINOv2 and a fixed

\tau_{1}=0.99

, isolating the effect of the training pipeline. We report target-domain PQ at different training checkpoints, averaged over three random seeds.

	$\tau_{1}=0.0$	$\tau_{1}=0.5$	$\tau_{1}=0.6$	$\tau_{1}=0.7$	$\tau_{1}=0.8$	$\tau_{1}=0.9$	$\tau_{1}=0.99$	Avg
	Synthia $\rightarrow$ Vistas
constant	–	$\text{40.3}_{\pm 2.8}$	$\text{42.2}_{\pm 1.6}$	$\text{42.4}_{\pm 4.3}$	$\text{37.2}_{\pm 2.6}$	$\text{42.3}_{\pm 1.7}$	$\text{44.4}_{\pm 1.2}$	$\text{41.5}_{\pm 2.4}$	$\Rsh$
per-class adaptive	$\text{44.8}_{\pm 0.1}$	$\text{43.6}_{\pm\textbf{0.4}}$	$\text{43.8}_{\pm\textbf{0.8}}$	$\text{44.6}_{\pm\textbf{1.0}}$	$\text{44.7}_{\pm\textbf{1.2}}$	$\text{44.2}_{\pm\textbf{0.6}}$	$\text{44.5}_{\pm\textbf{0.2}}$	$\text{44.3}_{\pm\textbf{0.4}}$	+2.8
	Cityscapes $\rightarrow$ ACDC
constant	–	$\text{53.6}_{\pm 2.6}$	$\text{53.5}_{\pm 2.0}$	$\text{53.4}_{\pm\textbf{0.5}}$	$\text{51.7}_{\pm 3.1}$	$\text{54.4}_{\pm 1.1}$	$\text{54.1}_{\pm 1.5}$	$\text{53.5}_{\pm 0.9}$	$\Rsh$
per-class adaptive	$\text{56.3}_{\pm 0.6}$	$\text{56.6}_{\pm\textbf{1.0}}$	$\text{56.6}_{\pm\textbf{0.6}}$	$\text{56.5}_{\pm\text{1.1}}$	$\text{56.2}_{\pm\textbf{0.3}}$	$\text{56.1}_{\pm\textbf{0.7}}$	$\text{57.0}_{\pm\textbf{0.6}}$	$\text{56.5}_{\pm\textbf{0.3}}$	+3.0

Table 11: Comparison of constant and per-class adaptive

\tau_{1}

for different starting values on Synthia

\rightarrow

Vistas (top) and Cityscapes

\rightarrow

ACDC (bottom), using single-stage MC-PanDA with DINOv2-Base. For each starting value, we report mean

{}_{\pm\text{std}}

panoptic quality (PQ) over three random seeds. The last column shows the average across all starting values.

Figure 9: Per-class PQ performance (left) and evolution of the adaptive confidence threshold

\tau_{1}

during training, alongside the constant baseline

\tau_{1}=0.99

(middle), for Cityscapes

\rightarrow

ACDC. Per-class adaptive

\tau_{1}

can accelerate learning for rare classes such as terrain and wall , while also enabling additional late-stage improvements for frequent classes such as car .

$f$	Synthia $\rightarrow$ Vistas	City $\rightarrow$ ACDC
mean	$\text{44.4}_{\pm 0.3}$	$\text{54.3}_{\pm 1.7}$
median	$\text{44.8}_{\pm 1.3}$	$\text{55.2}_{\pm 0.7}$
3. quartile	$\text{44.8}_{\pm 0.1}$	$\text{56.3}_{\pm 0.6}$

Table 12: Different aggregation functions

f

for per-class adaptive

\tau_{1}

. We report mean

{}_{\pm\text{std}}

panoptic quality (PQ) over three random seeds.

Outer aggregation	Syn $\rightarrow$ Vistas	City $\rightarrow$ ACDC	City $\rightarrow$ MUSES
mean	$\text{43.5}_{\pm 1.7}$	$\text{54.6}_{\pm 0.1}$	$\text{52.0}_{\pm 0.6}$
3. quartile	$\text{43.9}_{\pm 0.0}$	$\text{56.3}_{\pm 0.7}$	$\text{52.0}_{\pm 1.2}$
max	$\text{44.8}_{\pm 0.1}$	$\text{56.3}_{\pm 0.6}$	$\text{52.4}_{\pm 0.4}$

Table N2: Effect of the cross-instance aggregation over teacher-predicted masks of the same class in Eq. ( 5 ). We keep the instance-wise aggregation

f

fixed to the third quartile and vary the outer aggregation over instances. We report mean

\pm

std PQ over three random seeds.

$\tau_{1}$	Cityscapes $\rightarrow$ ACDC ( $\text{PQ}_{19}$ )
$\tau_{1}$	$\tau_{1}=0.0$	$\tau_{1}=0.5$	$\tau_{1}=0.99$
global adaptive	$\text{56.1}_{\pm 0.7}$	$\text{55.3}_{\pm 0.5}$	$\text{55.4}_{\pm 1.0}$
per-class adaptive	$\text{56.3}_{\pm 0.6}$	$\text{56.6}_{\pm 1.0}$	$\text{57.0}_{\pm 0.6}$

Table 13: Comparison of global-adaptive and per-class adaptive

\tau_{1}

on Cityscapes

\rightarrow

ACDC for different initial threshold values. We report mean

{}_{\pm\text{std}}

panoptic quality (PQ) over three random seeds.

	Synthia $\rightarrow$ Vistas ( $\text{PQ}_{16}$ )							Avg
$\tau_{1}{=}$	0.0	0.5	0.6	0.7	0.8	0.9	0.99	Avg
$\tau_{2}{=}0.7$	45.2	44.5	44.5	45.1	45.3	44.8	45.2	$\text{44.9}_{\pm 0.3}$
$\tau_{2}{=}0.8^{\dagger}$	44.8	43.6	43.8	44.6	44.7	44.2	44.5	$\text{44.3}_{\pm 0.4}$
$\tau_{2}{=}0.9$	44.4	45.2	45.0	44.5	44.8	44.8	43.8	$\text{44.6}_{\pm 0.4}$
	Cityscapes $\rightarrow$ ACDC ( $\text{PQ}_{19}$ )							Avg
$\tau_{1}{=}$	0.0	0.5	0.6	0.7	0.8	0.9	0.99	Avg

Table 14: Robustness of confidence-based point filtering (CBPF) with respect to the threshold

\tau_{2}

on: Synthia

\rightarrow

Vistas, Cityscapes

\rightarrow

ACDC, and Cityscapes

\rightarrow

MUSES, evaluated across different initial values of the per-class adaptive

\tau_{1}

\dagger

marks the default setting used in MC-PanDA ++ ,

\tau_{2}=0.8

Method

Backbone

Synthia (

\text{PQ}_{16}

)

UrbanSyn (

\text{PQ}_{19}

)

Cityscapes (

\text{PQ}_{19}

)

Cityscapes

Vistas

Cityscapes

Vistas

ACDC

MUSES

Source-only

DINOv2-Base

36.0

\Rsh

33.7

\Rsh

48.3

\Rsh

42.1

\Rsh

48.7

\Rsh

42.0

\Rsh

MC-PanDA ++

DINOv2-Base

49.6

+13.6

44.8

+11.1

57.2

+8.9

49.5

+7.4

56.3

+7.6

52.4

+10.4

Source-only

DINOv2-Large

39.9

\Rsh

37.9

\Rsh

51.1

\Rsh

45.2

\Rsh

52.3

\Rsh

46.4

\Rsh

MC-PanDA ++

DINOv2-Large

54.2

+14.3

47.6

+9.7

59.1

+8.0

52.7

+7.5

60.1

+7.8

55.6

+9.2

Table 15: Overview of the final MC-PanDA ++ results across six benchmarks. Improvements over the source-only fine-tuning baseline are shown for reference. All MC-PanDA ++ results are reported as mean

{}_{\pm\text{std}}

over three random seeds.

Panel of Figure 10 — Figure 10: Illustration of the MLS factor $\lambda_{i}$ for two model checkpoints during training, shown here for masks of classes road and traffic sign . As training progresses, the predicted masks become more complete and visually coherent, and the corresponding $\lambda_{i}$ values increase accordingly. A strong correlation is observed between the value of $\lambda_{i}$ and the visual quality of the predicted mask.

Panel of Figure 11 — Figure 11: Visual example illustrating the complementary roles of Mask-wide Loss Scaling (MLS) and Confidence-based Point Filtering (CBPF) in selecting reliable pixels for domain-adaptive learning, as also illustrated in MC-PanDA ( Martinović et al., 2024 ) . The middle image highlights two regions where back-propagation is suppressed due to low sampling affinity. The left and right images show that these regions correspond to false-negative predictions in the person and motorcycle masks, respectively.

Panel of Figure 12 — Figure 12: Predictions comparison on a Vistas image between source-only models trained on Synthia (middle left) and UrbanSyn (bottom left), and MC-PanDA ++ (right). Switching the synthetic source from Synthia to UrbanSyn already yields more coherent predictions, even without adaptation. MC-PanDA ++ further improves segmentation, particularly for distant cars , wall / fence regions, and small traffic signs . Best viewed zoomed in.

Panel of Figure 13 — Figure 13: Qualitative comparison of MC-PanDA ++ against a source-only model (trained solely on Cityscapes) on the clear-to-adverse benchmarks City $\rightarrow$ ACDC (top) and City $\rightarrow$ MUSES (bottom). MC-PanDA ++ yields substantially more accurate segmentations, especially in challenging regions such as person instances in top row and distant cars in bottom row, despite being trained only on clear-weather daytime ground-truth annotations. Best viewed zoomed in.

Panel of Figure 14 — Figure 14: Representative failure cases of MC-PanDA ++ . Left: UrbanSyn $\rightarrow$ Vistas example illustrating source-domain coverage limitations, where target-domain regions such as sea and boats are mapped to the closest available road-scene categories, such as road and car . Right: Cityscapes $\rightarrow$ ACDC nighttime example, where distant or weakly visible objects, such as train and bicycle regions, remain difficult to segment under limited visual evidence.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 32 — Figure A1: Example of a mislabeled UrbanSyn image (ID 7014 ) where building , vegetation , and traffic sign regions are incorrectly annotated as road or sky .

UrbanSyn filtering	USyn $\rightarrow$ City	USyn $\rightarrow$ Vistas
–	$\text{57.0}_{\pm 0.7}$	$\text{49.2}_{\pm 0.5}$
✓	$\text{57.2}_{\pm 0.7}$	$\text{49.5}_{\pm 0.2}$

Appendix

Table A1: Impact of UrbanSyn filtering on the final performance of MC-PanDA ++ . We report mean

\pm

std PQ over three random seeds.

Panel of Figure 34 — Figure A2: Illustration of SegMix during training. For each target image, half of the segments from a labeled source image are randomly selected and pasted onto the target image. This differs from ClassMix, which selects half of the classes and pastes all associated segments.

Panel of Figure 35 — Figure A3: Visual comparison of per-mask and all-mask point filtering ( cf. Table 5 main manuscript). The top row shows a person mask (left), uncertain pixels retained by all-mask filtering (middle), and those retained by per-mask filtering (right). The bottom row presents the same intermediate outputs for a motorcycle mask. As originally reported in the supplement of MC-PanDA ( Martinović et al., 2024 ) , all-mask filtering more reliably identifies false negative pixel assignments than per-mask filtering, a behavior that complements and explains the quantitative differences observed in Tab. 5 .

Panel of Figure 36 — Figure A4: Additional failure cases of MC-PanDA ++ . Left: UrbanSyn $\rightarrow$ Vistas example illustrating source-domain coverage limitations, where target-domain snow regions are mapped to the closest available road-scene categories, such as road and sidewalk . Right: UrbanSyn $\rightarrow$ Vistas example illustrating an annotation-policy ambiguity, where person depictions on billboards are segmented as person instances.

Group	Class	# pairs	Spearman $\rho$
Stuff	road	1599	0.52
	sidewalk	1400	0.65
	building	1420	0.87
	wall	848	0.66
	fence	700	0.59
	pole	1587	0.65

Appendix

Table A2: Full per-class Spearman

\rho

correlation between the MLS weight

\lambda_{i}

and pseudo-mask quality measured by IoU on Cityscapes

\rightarrow

ACDC. For thing classes, correlations are computed over matched prediction–ground-truth instance pairs.

Method	Encoder / setting	Synthia $\rightarrow$ City	Synthia $\rightarrow$ Vistas
EDAPS ( Saha et al., 2023 )	MiT-B5	41.2	36.6
LIDAPS ( Mansour et al., 2025 )	MiT-B5	44.8	38.0
EDAPS ‡ ( Saha et al., 2023 )	DINOv2-B, $512{\times}512$ crop	36.1	38.9
EDAPS ‡ ( Saha et al., 2023 )	DINOv2-B, $512{\times}1024$ crop	42.0	40.5
LIDAPS ‡ ( Mansour et al., 2025 )	DINOv2-B, $512{\times}512$ crop	40.2	37.0
LIDAPS ‡ ( Mansour et al., 2025 )	DINOv2-B, $512{\times}1024$ crop	44.7	35.8

Appendix

Table A3: Extended performance (PQ 16 ) comparison with EDAPS and LIDAPS on Synthia

\rightarrow

Cityscapes and Synthia

\rightarrow

Vistas. ‡ denotes our DINOv2-B reimplementations using the official codebases, a ViTDet-style feature pyramid, and GroupNorm normalization. For MC-PanDA ++ , we report the mean over three random seeds.

		Training					Inference ( $1024{\times}2048$ )
Method	Backbone	Iters.	Crop	s/iter $\downarrow$	GPU-h $\downarrow$	Peak mem. $\downarrow$	FPS $\uparrow$	Peak mem. $\downarrow$	PQ $\uparrow$
LIDAPS	MiT-B5	$40k{+}10k$	$512{\times}512$	1.77	24.5	17.5 GiB	2.1	6.6 GiB	44.8
LIDAPS ‡	DINOv2-B	$40k{+}10k$	$512{\times}1024$	1.83	25.4	29.2 GiB	2.4	8.2 GiB	44.7
MC-PanDA ++ (85k ckpt.)	DINOv2-B	$85k$	$512{\times}1024$	1.04	24.6	24.7 GiB	3.4	4.0 GiB	48.6
MC-PanDA ++ (final)	DINOv2-B	$110k$	$512{\times}1024$	1.04	31.9	24.7 GiB	3.4	4.0 GiB	49.6

Appendix

Table A4: Compute, memory, and inference-time comparison with LIDAPS on Synthia

\rightarrow

Cityscapes. For LIDAPS, training includes the 40k main stage and the 10k instance-mixing stage. Inference is measured on 500 Cityscapes images (

1024{\times}2048

), reporting the mean runtime with data loading excluded. ‡ denotes our DINOv2-B reimplementation of LIDAPS, described in Appendix F .

initial $\tau_{1}{=}$	0.0	0.5	0.6	0.7	0.8	0.9	0.99	Avg.
UrbanSyn $\rightarrow$ Vistas	49.5	49.2	49.7	49.7	49.3	49.4	49.5	$\text{49.5}_{\pm 0.2}$

Appendix

Table A5: Robustness of adaptive class-dependent MLS to initial values of

\tau_{1}

on UrbanSyn

\rightarrow

Vistas. Each value is averaged over three random seeds. The last column reports the mean and standard deviation across initial

\tau_{1}

values.

Five most frequent classes

pole (99.1%)

road (98.6%)

building (98.6%)

vegetation (97.2%)

car (95.2%)

Avg.

fixed

\tau_{1}=0.99

42.2

\Rsh

94.9

\Rsh

79.6

\Rsh

74.5

\Rsh

64.1

\Rsh

71.0

\Rsh

adaptive

\tau_{1}

44.8

+2.7

95.2

+0.3

79.8

+0.1

72.3

-2.3

71.4

+7.3

72.7

+1.6

Five rarest classes

train (4.8%)

bus (9.2%)

truck (12.1%)

motorcycle (17.2%)

wall (32.6%)

Avg.

fixed

\tau_{1}=0.99

60.0

\Rsh

62.0

\Rsh

40.8

\Rsh

31.0

\Rsh

48.9

\Rsh

48.5

\Rsh

Appendix

Table A6: Effect of per-class adaptive thresholding on the five most frequent and five rarest classes for Cityscapes

\rightarrow

ACDC. We compare the fixed threshold

\tau_{1}=0.99

with our per-class adaptive threshold. Class frequency, shown in parentheses, is measured as the percentage of source-domain training images in which the class appears at least once. Results are per-class PQ averaged over three random seeds.

	road	sidewalk	building	wall	fence	pole	tr. light	tr. sign	vegetation	sky	person	rider	car	bus	motorcycle	bicycle	$\text{PQ}_{16}$
Method	Synthia $\rightarrow$ Cityscapes
CVRN	86.6	33.8	74.6	3.4	0.0	10.0	5.7	13.5	80.3	76.3	26.0	18.0	34.1	37.4	7.3	6.2	32.1
UniDAF	73.7	26.5	71.9	1.0	0.0	7.6	9.9	12.4	81.4	77.4	27.4	23.1	47.0	40.9	12.6	15.4	33.0
UniDAF-PSN	87.7	34.0	73.2	1.3	0.0	8.1	9.9	6.7	78.2	74.0	37.6	25.3	40.7	37.4	15.0	18.8	34.2
EDAPS	77.5	36.9	80.1	17.2	1.8	29.2	33.5	40.9	82.6	80.4	43.5	33.8	45.6	35.6	18.0	2.8	41.2
LIDAPS	80.8	48.8	80.8	17.6	2.5	29.9	34.6	42.9	82.8	82.9	44.4	40.5	51.7	39.2	27.4	10.7	44.8

Appendix

Table A7: Per-class PQ performance evaluation on four standard domain adaptation setups and comparison with the state of the art. All experiments for MC-PanDA and MC-PanDA ++ are averaged over three random seeds.

Explore similar work

Apr 21, 2026cs.CV

PanDA: Unsupervised Domain Adaptation for Multimodal 3D Panoptic Segmentation in Autonomous Driving

This paper presents the first study on Unsupervised Domain Adaptation (UDA) for multimodal 3D panoptic segmentation (mm-3DPS), aiming to improve generalization under domain shifts commonly encountered in real-world autonomous driving. A straightforward solution is to employ a pseudo-labeling strategy, which is widely used in UDA to generate supervision for unlabeled target data, combined with an mm-3DPS backbone. However, existing supervised mm-3DPS methods rely heavily on strong cross-modal complementarity between LiDAR and RGB inputs, making them fragile under domain shifts where one modality degrades (e.g., poor lighting or adverse weather). Moreover, conventional pseudo-labeling typically retains only high-confidence regions, leading to fragmented masks and incomplete object supervision, which are issues particularly detrimental to panoptic segmentation. To address these challenges, we propose PanDA, the first UDA framework specifically designed for multimodal 3D panoptic segmentation. To improve robustness against single-sensor degradation, we introduce an asymmetric multimodal augmentation that selectively drops regions to simulate domain shifts and improve robust representation learning. To enhance pseudo-label completeness and reliability, we further develop a dual-expert pseudo-label refinement module that extracts domain-invariant priors from both 2D and 3D modalities. Extensive experiments across diverse domain shifts, spanning time, weather, location, and sensor variations, significantly surpass state-of-the-art UDA baselines for 3D semantic segmentation.

Yining Pan, Shijie Li, Yuchen Wu +2

Singapore University of Technology and Design · Institute for Infocomm Research (I2R), A*STAR, Singapore

Panoptic Segmentation Autonomous Driving Domain Adaptation High Confidence Pseudo-Labels Cross-Modal

Ask PDF

May 5, 2026cs.CV

Dual-Foundation Models for Unsupervised Domain Adaptation

Semantic segmentation provides pixel-level scene understanding essential for autonomous driving and fine-grained perception tasks. However, training segmentation models requires costly, labor-intensive annotations on real-world datasets. Unsupervised Domain Adaptation (UDA) addresses this by training models on labeled synthetic data and adapting them to unlabeled real images. While conceptually simple, adaptation is challenging due to the domain gap, i.e., differences in visual appearance and scene structure between synthetic and real data. Prior approaches bridge this gap through pixel-level mixing or feature-level contrastive learning. Yet, these techniques suffer from two major limitations: (1) reliance on high-confidence pseudo-labels restricts learning to a subset of the target domain, and (2) prototype-based contrastive methods initialize class prototypes from source-trained models, yielding biased and unstable anchors during adaptation. To address these issues, we propose a dual-foundation UDA framework that leverages two complementary foundation models. First, we employ the Segment Anything Model (SAM) with superpixel-guided prompting to enable learning from a broader range of target pixels beyond high-confidence predictions. Second, we incorporate DINOv3 to construct stable, domain-invariant class prototypes through its robust representation learning. Our method achieves consistent improvements of +1.3% and +1.4% mIoU over strong UDA baselines on GTA-to-Cityscapes and SYNTHIA-to-Cityscapes, respectively.

Yerin Cheon, Aruna Balasubramanian, Francois Rameau

Stony Brook University, Stony Brook, NY, USA · SUNY Korea, The State University of New York, Korea

Domain Adaptation Semantic Segmentation Recent Vision Foundation Models Wireless Foundation Models Fine-Grained Perception Pixels

Ask PDF

Jul 11, 2026cs.CV

CVKD-UDA: Cross-View Knowledge Distillation for 3D Unsupervised Domain Adaptive Segmentation

3D unsupervised domain adaptive (UDA) segmentation mitigates the high cost of manual annotations of the new domain data. Self-training has emerged as the dominant approach in this area, where its success heavily depends on a well-initialized warm-up model to generate reliable pseudo labels. However, existing methods often depend on source supervision or output-level adversarial alignment to obtain the warm-up model, which suffer from limited generalization and training instability due to the large domain gap between domains. Constructing domain-similar representations is an effective way to bridge this gap. In this work, we propose CVKD-UDA, which revisits voxel size as a core design factor to construct domain-similar representations and leverages cross-view complementary cues to balance transferability and discriminability of the warm-up model. First, we generate two complementary views by varying voxel sizes and introduce a cross-view knowledge distillation (CVKD) to enhance generalization and target perception of the model. Second, to balance transferability and discriminability, we design a lightweight Decouple-Adapter and an auxiliary imitation classifier to decouple cross-view knowledge transfer. Extensive experiments on two benchmarks demonstrate that CVKD-UDA effectively improves the performance of self-training methods and provides a new perspective for 3D UDA segmentation. Our code will be available at GitHub.

Zhimin Yuan, Ming Cheng, Shangshu Yu +4

School of Artificial Intelligence, Nanyang Normal University, Nanyang 473061, China · Fujian Key Laboratory of Sensing and Computing for Smart Cities, School of Informatics, Xiamen University, Xiamen 361005, China · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China

Interactive 3D Segmentation Domain Adaptation U-Net Knowledge Distillation

Ask PDF