PaperlayerView exact revision on arXiv

Source-linked AI summary

Unsupervised Domain Adaptation in Semantic Segmentation: a Review

Marco Toldo, Andrea Maracani, Umberto Michieli, Pietro Zanuttigh

arXiv:2005.10876v1cs.CVcs.LGeess.IV

TL;DR

Semantic segmentation depends on large labeled datasets, which may be unavailable or costly to annotate. This survey organizes UDA approaches by adaptation level and method category. Its autonomous-driving comparison reports variation across methods and evaluation protocols, with ResNet-based approaches outperforming competitors.

  • Problem

    Semantic segmentation requires large labeled training sets that may be unavailable, while UDA approaches for the task have proliferated and need to be organized.

    Source: §5 · p. 28

  • Method

    The survey groups UDA methods by adaptation level and reviews seven non-mutually-exclusive categories of approaches.

    Source: §5 · p. 28

  • Results

    In GTA5-to-Cityscapes comparisons, mIoU varies greatly with method and evaluation protocol, while ResNet-based approaches outperform competitors.

    Source: §4.3 · p. 28

  • Takeaways & Limitations

    The survey identifies UDA for semantic segmentation as an open field, with open-set and boundless-set adaptation among future research directions.

    Source: §5 · p. 29

    Even the best approaches remain far from supervised training on the target dataset.

    Source: §5 · p. 29

Abstract

from arXiv · show

The aim of this paper is to give an overview of the recent advancements in the Unsupervised Domain Adaptation (UDA) of deep networks for semantic segmentation. This task is attracting a wide interest, since semantic segmentation models require a huge amount of labeled data and the lack of data fitting specific requirements is the main limitation in the deployment of these techniques. This problem has been recently explored and has rapidly grown with a large number of ad-hoc approaches. This motivates us to build a comprehensive overview of the proposed methodologies and to provide a clear categorization. In this paper, we start by introducing the problem, its formulation and the various scenarios that can be considered. Then, we introduce the different levels at which adaptation strategies may be applied: namely, at the input (image) level, at the internal features representation and at the output level. Furthermore, we present a detailed overview of the literature in the field, dividing previous methods based on the following (non mutually exclusive) categories: adversarial learning, generative-based, analysis of the classifier discrepancies, self-teaching, entropy minimization, curriculum learning and multi-task learning. Novel research directions are also briefly introduced to give a hint of interesting open problems in the field. Finally, a comparison of the performance of the various methods in the widely used autonomous driving scenario is presented.

1. Introduction

Deep semantic segmentation depends on large, scenario-matched labeled datasets, but pixel-level annotation is expensive. This review organizes research on adapting models to related target domains without target labels.

1.1. Semantic Segmentation

  • Semantic segmentation assigns semantic labels to every pixel, making it a dense, challenging task whose deep models depend on massive labeled datasets; recent state-of-the-art approaches use encoder-decoder architectures.Source: §1.1 · p. 2

1.2. Domain Adaptation (DA)

  • Domain adaptation addresses the distribution shift between source and target domains, whose relatedness affects performance; selecting a suitable source is key to extracting useful clues.Source: §1.2 · p. 3

1.3. Unsupervised Domain Adaptation (UDA)

  • UDA adapts a model trained on labeled source samples using unlabeled target samples; target labels are unavailable, and the standard setting assumes the same target classes.Source-only training can perform poorly even when the visual domain shift is small.Source: §1.3 · p. 3

1.4. Application Motivations

  • UDA addresses application-specific data scarcity when generic public labels transfer poorly; synthetic-to-real adaptation is especially prominent in autonomous driving and also relevant to industrial applications.Source: §1.4 · p. 4

1.5. Outline

  • The survey formulates UDA stages, categorizes alignment methods after precursor weak-supervision work, reviews synthetic-to-real road-scene results by architecture and scenario, and outlines future directions.Source: §1.5 · pp. 4–5

2. Unsupervised Domain Adaptation for Semantic Segmentation

Unsupervised domain adaptation learns a segmentation function from labeled source data and unlabeled target data. It addresses source–target distribution shift through adaptation at the image, feature, output, or other network levels.

2.1. Problem Formulation

  • Unsupervised DA uses labeled source samples and completely unlabeled target samples to learn a hypothesis that performs well on the target domain.DA is supervised when target labels are available for all samples and semi-supervised when they are available for only some.Source: §2.1 · p. 5
  • Category configurations distinguish closed-set (CS = CT), partial (CS ⊃ CT), open-set (CS ⊂ CT), open-partial (CS ≠ CT with overlap), and boundless DA (CS ⊂ CT, CL = CS ∪ CT).Open-set target-only categories are usually learned as an unknown class, whereas Boundless DA learns them individually [23].Source: §2.1 · p. 5

2.2. UDA in Semantic Segmentation: Adaptation Spaces

  • Because covariate shift harms target predictions, UDA primarily seeks to bridge source and target distributions to lessen the performance drop.The review presents adaptation levels as a framework for categorizing methods.Source: §2.2 · p. 6
  • Input-level adaptation translates source images toward target appearance, or target images toward source, but image-marginal matching alone may lose semantic consistency.Approaches use reconstruction constraints, prediction uniformity, or other techniques to preserve content during translation.Source: §2.2 · pp. 6–7
  • Feature-level methods align latent representations globally or class-wise so a source-trained classifier can segment both domains, but simple alignment is limited by feature complexity and absent joint-label alignment [24].Feature adaptation is therefore combined with complementary techniques or specific arrangements.Source: §2.2 · p. 7
  • Output-level adaptation aligns segmentation maps or final-layer class outputs, enabling adversarial matching and use of target label statistics or source label-distribution priors.Target label statistics can provide self-constructed weak supervision; source priors can capture high-level structural properties.Source: §2.2 · pp. 7–8
  • Ad-hoc network-level methods align activations at intermediate levels to capture high-level patterns, with gradients flowing through the segmentation network beyond the end of feature extraction.Source: §2.2 · p. 8

3. Review of Unsupervised Domain Adaptation strategies

The review introduces weakly and semi-supervised methods as early domain-adaptation approaches, then organizes UDA methods into seven categories and identifies future research directions. It explains how the categories address domain alignment and target-domain learning through distinct mechanisms.

Domain adversarial discriminative approaches

  • Domain-adversarial discriminative approaches use adversarial learning to produce data with a statistical distribution similar to the training samples.Source: §3 · p. 9

Generative-based approaches

  • Generative-based approaches translate data between domains to create target-like training data or representations closer to target-domain characteristics.Source: §3 · p. 9

Classifier discrepancy approaches

  • Classifier-discrepancy approaches use multiple dense classifiers on one encoder to identify less-adapted target representations and encourage cross-domain alignment far from decision boundaries.They use an adversarial-like strategy.Source: §3 · p. 9

Self-training approaches

  • Self-training approaches use confidence-based selection of reliable pseudo-labels from current predictions to guide and self-supervise learning.Source: §3 · p. 9

Entropy minimization methods

  • Entropy-minimization methods minimize target output-map entropy to promote source-like overconfidence and well-clustered target feature representations.Source: §3 · p. 9

Curriculum learning approaches

  • Curriculum-learning approaches first tackle easier tasks to infer target-domain properties, such as global label distributions, then train predictions to follow those properties.Source: §3 · p. 9

Multi-tasking methods

  • Multi-tasking methods solve multiple tasks simultaneously to improve the extraction of invariant feature representations.Source: §3 · p. 9

Recent research directions

  • The review concludes by identifying recent research directions that merit further development.Source: §3 · p. 9

3.1. Weakly- and Semi- Supervised Learning

Weakly and partially supervised methods preceded unsupervised domain adaptation, addressing the high cost of pixel-level labels by using weaker annotations or combining them with a smaller set of pixel-labeled samples. The section surveys approaches that infer, constrain, or iteratively refine pixel-level segmentation labels.

3.1. Weakly- and Semi- Supervised Learning

  • Weak supervision used MIL-based STF label inference and [26], EM label estimation, CCNN output constraints [28], and staged simple-to-complex training.The simple-to-complex method first derives maps from weakly annotated simple images, then trains three networks sequentially to segment more complex images.Source: §3.1 · pp. 9–10
  • The semi-supervised architecture [30] combines a classification network, bridging layers, and a segmentation network, enabling training with few pixel-annotated samples.Weakly trained classification features yield class-specific activation maps that bridging layers refine for the segmentation network, reducing its parameter count and simplifying its task.Source: §3.1 · p. 10
  • Bounding-box supervision [31] iteratively selects candidate masks that overlap boxes, trains a segmentation network on those labels, and feeds its outputs back to improve label selection.The candidate masks are generated by region proposal methods and fixed during training; selected labels and network outputs improve together across iterations.Source: §3.1 · p. 10
  • A GAN approach [32] makes the discriminator assign each pixel a semantic-class or fake label, while another approach incorporates weak image-level annotations into conditional GAN inputs.The discriminator is trained using generated data, unlabeled data for regularization, and pixel-labeled data.Source: §3.1 · p. 10
  • Self-supervised methods use inferred pixel activations as pseudo-ground truth [33], including CAM seed expansion with SRG, progressive adversarial erasing [35], and other self-learning techniques [36], [37], and [38].SRG expands sparse discriminative regions to neighboring pixels with similar features, and segmentation outputs help update those labels iteratively.Source: §3.1 · p. 10
  • A four-stage cross-task and cross-domain framework [39] transfers synthetic-to-real segmentation by learning shared features and task mappings, using easier-to-acquire depth maps as weak annotations.It trains a shared representation across domains for one task, then learns the second task and a mapping between task features on the first domain before transferring to the second.Source: §3.1 · pp. 10–11

3.2. Domain Adversarial Discriminative

Adversarial domain adaptation uses a discriminator to align source and target representations while source supervision trains the segmentation network. Methods apply this alignment to internal features or segmentation outputs, sometimes adding class-wise or structured prediction signals.

Feature Adversarial Adaptation

  • Adversarial adaptation repurposes the GAN discriminator as a source–target domain classifier [41] [43], aligning representations while source supervision trains semantic segmentation.The discriminator learns a measure of domain discrepancy and guides the predictor toward domain invariance. Hoffman et al. [44] first applied this approach to semantic segmentation, using intermediate features for global alignment and image-level label-distribution constraints for category-specific alignment.Source: §3.2 · p. 11
  • Feature-level systems augment global alignment with spatial structure, multi-scale representations, information filtering, or image translation.Chen et al. [45] match target activations to an ImageNet-pretrained network and align spatial regions; Zhang et al. [46] add ASPP; Luo et al. [51] use SIB to preserve domain-invariant discriminative cues and a module to aid less frequent classes. Other methods [47] [49] [50] [52] combine source-to-target image translation with adversarial alignment of latent embeddings.Source: §3.2 · p. 12
  • Class-wise adversarial methods complement global alignment to reduce cross-class negative transfer and address class-boundary and class-frequency imbalances.Chen et al. use one discriminator per class and grid-level soft pseudo labels; Du et al. add discriminator independence and adaptively re-weight class contributions to address boundary errors and dominance by more readily adapted classes.Source: §3.2 · p. 12
  • Reconstruction-based methods align domains by adversarially matching reconstructed images, training features to generate target-like or source-like images across domains.The encoder competes with the domain discriminator using reconstructed images, encouraging source and target features to be projected back into either image space.Source: §3.2 · p. 12

Output Adversarial Adaptation

  • Output-level methods align source and target prediction maps in a lower-dimensional space expected to retain semantic information.Tsai et al. use discriminators at multiple depths; Chen et al. [60] jointly align segmentation and depth maps, while Luo et al. use confidence-based co-training to focus on less-adapted regions. Methods [66] instead make source and target predictions resemble source ground-truth maps, producing indirect alignment.Source: §3.2 · pp. 12–13
  • Other output-space approaches align self-information maps or clustered patches to expose confidence and structured prediction patterns to the discriminator.Vu et al. align self-information statistics, encouraging confident, low-entropy target predictions and pushing decision boundaries away from high-density regions. Tsai et al. use source-supervised patch clustering followed by adversarial alignment of clustered source and target representations.Source: §3.2 · p. 13

3.3. Generative-Based Approaches

Generative-based UDA adapts image appearance or style to reduce source–target visual differences while preserving source semantics and annotations for segmentation training. The section covers bidirectional and lighter-weight translation, semantic-consistency strategies, reverse translation, and style transfer.

3.3. Generative-Based Approaches

  • Unpaired image translation learns cross-domain mappings from unpaired data, but its ill-posedness means constraints are needed to obtain acceptable solutions.For segmentation adaptation, translated source images aim to have target-like statistics while retaining source semantic information and annotations.Source: §3.3 · pp. 13–14
  • CycleGAN uses two adversarial translation directions and cycle consistency to preserve scene geometry, but does not guarantee semantic consistency.Many works [47] [49] [50] [52] [70] [73] use this framework for input-level adaptation.Source: §3.3 · p. 14
  • Semantic-aware approaches [47] [49] [50] [52] [70] constrain translations using segmentation predictions, while Li et al. [47] alternately train translation and segmentation for mutual feedback and progressive adaptation.Because target predictions can be inaccurate without annotations, some works jointly optimize the generative and discriminative components [52] or use separate source and target predictors [50].Source: §3.3 · p. 14
  • Other CycleGAN adaptations preserve semantics through boundary-sensitive gradients, Fourier phase consistency [73], or intermediate domains intended to ease adaptation and improve generalization.Li et al. also use a semantic-aware discriminator to assess resemblance between original and translated samples.Source: §3.3 · pp. 14–15
  • To reduce CycleGAN’s computational burden, one-way generative methods [60] use related-task consistency, feature normalization, semantic regularization, or adversarial feature translation.These approaches omit the backward translation branch; Hong et al. model target-like feature representations without assuming a shared domain-invariant latent space.Source: §3.3 · p. 15
  • Yang et al. [63] translate target images toward the source, train with pseudo-labels, and use label-driven image reconstruction to guide category-wise feature alignment.Training directly on the source retains its original annotations and avoids semantic alterations that may occur in source-to-target translation.Source: §3.3 · p. 15
  • Style-transfer methods combine target style with source content, using latent decomposition, neural style transfer [46], feature-statistic re-normalization, photo-realistic transfer [83], or frequency replacement [24].MUNIT-based multi-modal translation transfers multiple target styles to one source image to increase training-data diversity; frequency replacement [24] needs no additional learnable module.Source: §3.3 · pp. 15–16

3.4. Classifier Discrepancy

The section presents classifier-discrepancy adaptation as a response to semantic-unaware feature alignment, surveying methods that use prediction disagreement to identify or correct poorly adapted target features.

3.4. Classifier Discrepancy

  • Feature-level adversarial critics align marginal distributions but ignore category-level joint distributions and decision boundaries, risking misclassification and negative transfer.Source: §3.4 · p. 16
  • ADR perturbs the classifier with dropout, then minimizes disagreement between two output maps to push target features away from decision boundaries.The classifier maximizes output variation to detect less-adapted features, while the encoder promotes categorical certainty and features diverse from decision boundaries.Source: §3.4 · pp. 16–17
  • Saito et al. replace ADR’s dropout with two source-supervised decoders whose maximized discrepancy reveals target support, reducing noise sensitivity but requiring an extra training stage.Source: §3.4 · p. 17
  • Luo et al. weight adversarial adaptation by two-classifier discrepancy, while Lee et al. use virtual-dropout masks and a unified objective to push dense regions and decision boundaries apart.Source: §3.4 · p. 17

3.5. Self-Training

Self-training adapts segmentation models by using confident target predictions as pseudo-labels alongside source annotations, promoting cross-domain feature alignment while risking the propagation of errors. Approaches address this risk through confidence filtering, balancing, ensembling, or auxiliary guidance.

3.5. Self-Training

  • Self-training uses confident target predictions as pseudo-labels alongside source labels, encouraging cross-domain feature alignment but risking overconfident errors that iterative training can propagate.Most methods filter pseudo-labels to retain predictions deemed more likely to be correct.Source: §3.5 · pp. 17–18
  • Zou et al. alternate network training and pseudo-label estimation, using class-wise confidence thresholds and source-derived spatial priors; their later method uses continuous pseudo-labels and confidence regularization.The class-wise thresholds aim to balance contributions across categories, while continuous pseudo-labels are intended to reduce misleading overconfident supervision and promote smoother outputs.Source: §3.5 · p. 18
  • Pizzati et al. use a learnable confidence threshold to refine and weight continuous pseudo-labels computed batch-wise, balancing uncertain pixels while retaining a global view across training.This approach avoids slow dataset-wide offline processing.Source: §3.5 · p. 18
  • Adversarial-discriminator methods [66] use discriminator outputs to select reliable target predictions, with region growing and class-wise adaptive thresholding [66] refining selection.The adaptive thresholds select the same fraction of confident pixels per class based on batch-wise discriminator probabilities, adding inter-class flexibility and adaptation over time.Source: §3.5 · p. 18
  • Ensembling methods [24] [70] generate target pseudo-labels from combined predictions; Yang et al. [24] average multiple spectrum-adapted segmenters for more robust adaptation across self-training rounds.Chen et al. combine low- and high-level target predictions using an additional ASPP dense classification module.Source: §3.5 · pp. 18–19
  • Choi et al. train a student to follow an averaged-weight teacher’s target predictions, using input noise and dropout for consistency; Zhou et al. [70] filter unreliable teacher predictions with an uncertainty module.The teacher provides auxiliary pseudo-label-like guidance for supervised training on target data.Source: §3.5 · p. 19

3.6. Entropy Minimization

Entropy minimization adapts segmentation by encouraging confident, well-clustered target predictions, and later methods address limitations of pixel-level entropy objectives with structural alignment, modified losses, and complementary adaptation modules.

3.6. Entropy Minimization

  • Target entropy minimization encourages confident, well-clustered target representations by pushing decision boundaries away from high-density regions, addressing uncertainty beyond semantic boundaries.Target entropy maps are described as unstable and noisy, and reducing this uncertainty is intended to narrow the accuracy gap between domains.Source: §3.6 · p. 19
  • The pixel-level objective treats prediction units independently, while Vu et al. align source and target entropy maps globally and enforce source-derived class-wise priors.A domain discriminator captures global patterns in entropy maps to support a more semantically meaningful cross-domain match and limit class imbalance toward easy classes.Source: §3.6 · p. 19
  • Chen et al. address gradients dominated by confident predictions with maximum squares loss and target-prediction-based category weights instead of source class statistics.Their loss makes gradients grow linearly with input probability, while category-wise weighting addresses class imbalance using the target distribution inferred from prediction maps.Source: §3.6 · pp. 19–20
  • Yang et al. [24] add entropy minimization as a low-overhead regularizer, strengthened by other alignment modules and a penalty focused on less-adapted, high-entropy target regions.The combined modules shift UDA toward semi-supervised learning, where entropy minimization is more effective.Source: §3.6 · p. 20

3.7. Curriculum Learning

Curriculum learning first infers target-domain properties through easier tasks, then uses them to guide segmentation training. The reviewed methods estimate label distributions, progressively adapt across fog levels, or combine curriculum learning with self-training.

3.7. Curriculum Learning

  • Curriculum approaches train the network to reproduce inferred target-domain properties, unlike self-training, which uses target pseudo-labels as estimates of the desired annotations.The inferred properties differ from the labels for the segmentation task and are used to support training on that more challenging task.Source: §3.7 · p. 20
  • In and, adaptation matches target global pixel-label frequencies and local anchor-superpixel label frequencies through target-domain cross-entropy.A logistic regression model estimates the global distribution, while an SVM selects representative anchor superpixels; the final objective also includes source pixel-wise cross-entropy.Source: §3.7 · p. 20
  • CMAda and gradually adapts clear-weather segmentation to dense fog through intermediate fog levels, using lighter-fog real images and next-step synthetic fog for supervision.Synthetic fog has tunable density β; an AlexNet regressor trained on synthetic images estimates fog density in both synthetic and real foggy images.Source: §3.7 · pp. 20–21
  • PyCDA merges curriculum learning and self-training by treating self-training pseudo-labels as curriculum properties.It replaces superpixels with small squared regions and infers curriculum properties with the segmentation network, avoiding additional models such as SVMs or logistic regression.Source: §3.7 · p. 21

3.8. Multi-Tasking

Multi-task adaptation uses auxiliary depth, geometric, and boundary information to improve target-domain segmentation. The methods differ in whether auxiliary information regularizes training, enters segmentation features, guides image adaptation, or defines additional tasks.

3.8. Multi-Tasking

  • SPIGAN uses privileged source information such as depth as regularization, whereas DADA feeds predicted depth into segmentation features during adversarial domain alignment.SPIGAN translates source images toward target style and predicts privileged information on source and adapted images; DADA adds residual auxiliary blocks and combines self-information with depth maps.Source: §3.8 · p. 21
  • GIO-Ada uses semantic and depth maps to guide adversarial adaptation at both the input and output levels.Its transform network takes source images with semantic and depth maps to create target-like images; a task network predicts semantic and depth maps whose outputs are discriminated as real or adapted.Source: §3.8 · p. 22
  • For, maximum-classifier-discrepancy adaptation is augmented by RGB-depth input fusion or simultaneous segmentation, depth-regression, and boundary-detection tasks.The multi-task variant uses RGB inputs and separate networks after the feature generator to solve the three tasks.Source: §3.8 · p. 22

3.9. New Research Directions

The section identifies more realistic adaptation settings in which target data may contain unseen classes, alongside continual learning for changing domains or progressively refined label spaces. It outlines methods that recognize unknown categories rather than assuming a fixed shared class set.

3.9. New Research Directions

  • Standard UDA assumes source and target share a defined class set, but target images may contain unseen categories, motivating more realistic adaptation settings [103].Open-set adaptation addressed unknown target categories in image classification; later approaches relaxed prior assumptions about target classes.Source: §3.9 · p. 22
  • Universal Domain Adaptation [103] clusters target samples as source-class or unknown, then minimizes entropy to align domains without target-class supervision.The further goal is to learn unseen target categories individually rather than merely acknowledge them as unknown.Source: §3.9 · p. 22
  • Unsupervised Open Domain Recognition minimizes feature-similarity-weighted instance-matching discrepancy, then transfers known-class knowledge to unknown classes with a graph CNN.Source: §3.9 · p. 22
  • For segmentation, BUDA [23] uses known class names as minimal semantic prior and separates adaptation of shared classes from zero-shot transfer to unseen classes.Its entropy-minimization strategy avoids aligning unseen classes incorrectly; a generative model synthesizes visual features conditioned on class descriptors for zero-shot learning.Source: §3.9 · p. 23
  • Continual learning addresses domain distributions that change incrementally; class-incremental learning adds tasks while preserving prior capabilities, and coarse-to-fine learning reuses coarser semantic knowledge.The area has been explored for semantic segmentation [110] [111], following earlier work on image classification and object detection.Source: §3.9 · p. 23

4. A Case Study: Synthetic to Real Adaptation for Semantic Understanding of Road Scenes

This case study examines adapting labeled synthetic urban-scene data to real-world street scenes, where target labels are not needed for training but limited annotations support evaluation. It compares methods across common synthetic-to-real benchmarks while noting that differing experimental protocols complicate direct accuracy comparisons.

4. A Case Study: Synthetic to Real Adaptation for Semantic Understanding of Road Scenes

  • UDA uses labeled synthetic source images for supervised training, while real target-domain labels are unnecessary for adaptation but limited annotations are used for testing and sometimes validation.Synthetic-to-real urban-scene adaptation is motivated by inexpensive synthetic labels and the high expense, time, and error risk of real-world annotation.Source: §4 · p. 23

4.1. Source Domain: Synthetic Datasets of Urban Scenes

  • GTA5 provides 24,966 labeled 1914 × 1052 synthetic street images across 19 classes, while SYNTHIA-RAND-CITYSCAPES provides 9,400 labeled 1280 × 760 images with 13- or 16-class evaluation.GTA5 uses a commercial game renderer for realistic American-style streets; SYNTHIA-RAND-CITYSCAPES offers varied European-style scenes but lower visual quality.Source: §4.1 · p. 24

4.2. Target Domain: Real World Datasets of Urban Scenes

  • Cityscapes and Mapillary are common real-world targets, with target training images used without labels for adaptation and validation images used for testing.Cityscapes covers 50 European cities; Mapillary offers images from varied devices and locations worldwide and annotations are often mapped to Cityscapes classes.Source: §4.2 · p. 24

4.3. Methods Comparison

  • The comparison focuses on GTA5 and SYNTHIA as source datasets and Cityscapes and Mapillary as targets, while differing resolutions, batch sizes, backbones, and training settings can hinder direct comparison.Per-class IoU and mean IoU are the most common metrics.Source: §4.3 · pp. 25, 28
  • ResNet-based approaches generally outperform competitors for GTA5-to-Cityscapes adaptation, while results vary by method and evaluation protocol.ResNet-101 is the most widely used backbone and is identified as a strong choice for comparisons with existing work.Source: Table 1 · p. 25, Figure 6 · p. 26, §4.3 · p. 28
  • For SYNTHIA-to-Cityscapes adaptation, results are reported over 13 or 16 classes; ResNet-101 generally achieves higher results than the more popular VGG-16.The table and figure summarize methods by backbone, with Figure 7 displaying mIoU16 entries for backbones with at least three entries.Source: Figure 7 · p. 26, Table 2 · p. 27, §4.3 · p. 28
  • 41.9 mIoU is the reported highest result for GTA5-to-Mapillary adaptation, and the authors identify Mapillary’s greater variability as a reason for future study.The comparison is described as limited, covering works [66], with [66] reported as the highest-performing approach to the authors’ knowledge.Source: §4.3 · p. 28

5. Conclusion and Future Directions

The paper surveys UDA methods for semantic segmentation, organizing them by adaptation level and by seven non-mutually-exclusive approach categories. It identifies substantial room for improvement and future directions in open-set and boundless-set UDA and continual learning.

5. Conclusion and Future Directions

  • The survey organizes UDA methods by adaptation level and seven non-mutually-exclusive categories, summarizing key approaches and their contributions.The adaptation levels are input, feature, output, and ad-hoc network levels; categories include adversarial learning, generative approaches, classifier-discrepancy analysis, self-training, entropy minimization, curriculum learning, and multi-task learning.Source: §5 · p. 28
  • The survey concludes with a synthetic-to-real road-scene adaptation case study, comparing works by backbone architecture.The paper frames this task as relevant to autonomous driving and commonly used to evaluate UDA methods.Source: §5 · pp. 28–29
  • UDA for semantic segmentation remains open, with the best approaches still below supervised target-dataset performance; future directions include open-set and boundless-set UDA.The authors also suggest that UDA knowledge may benefit continual learning, where data distributions change multiple times.Source: §5 · p. 29

Related papers

Loading 2005.10876v1…