SYNTHIA: Synthetic Urban Scene Benchmark

• SYNTHIA is a synthetic urban dataset offering photo-realistic RGB images, pixel-wise semantic labels, and depth data under varied weather and illumination conditions.
• It is widely used to evaluate sim-to-real domain adaptation in semantic segmentation, employing metrics like mIoU with task-specific class protocols.
• The dataset also supports auxiliary supervision with privileged signals (depth and edge labels) and extends to panoramic and 3D point-cloud scenarios.

to=arxiv_search.search 彩神争霸苹果_json {"query":"SYNTHIA dataset semantic segmentation domain adaptation", "max_results": 10} to=arxiv.search 天天中彩票的_json {"query":"SYNTHIA dataset semantic segmentation domain adaptation", "max_results": 10} SYNTHIA is a synthetic urban-scene benchmark used extensively in computer vision as a labeled source domain for semantic segmentation, domain adaptation, video understanding, and related scene-perception tasks. In the literature summarized here, it appears most often as a driving-simulator dataset of urban scenes that provides synthetic RGB images, pixel-wise semantic labels, and, in several settings, depth / z-buffer information under varied weather and illumination conditions. Derivative benchmarks such as SYNTHIA-PANO and Omni-SYNTHIA extend it to panoramic and omnidirectional settings, while a separate group of papers reuses the same name for unrelated systems in microscopy, design generation, persona modeling, and solar physics (Lee et al., 2018, Xu et al., 2019, Zhang et al., 2019, Rahimzadeh et al., 20 Jul 2025).

1. Dataset composition and benchmark conventions

In the driving-vision literature, SYNTHIA is described as a synthetic dataset of urban scenes rendered in a 3D environment, and more specifically as a driving-simulator dataset of urban scenes with synthetic RGB images, pixel-wise semantic segmentation labels, depth / z-buffer information, and varied weather and illumination conditions. A commonly used subset is SYNTHIA-RAND-CITYSCAPES, reported as containing 9,400 images at resolution $1280 \times 760$, with labels consistent with Cityscapes; this is the standard source-domain configuration in much of the synthetic-to-real adaptation literature (Zhang et al., 2020).

Benchmark protocols differ materially across papers. In SYNTHIA $\rightarrow$ Cityscapes adaptation, performance is commonly reported as mIoU@16 and mIoU@13, because not all methods evaluate the same subset of classes, and one translation-based paper further notes that SYNTHIA lacks annotations for terrain, truck, and train, so evaluation is performed on the 16 common classes (Brehm et al., 2021). A latent-diffusion UDA paper describes SYNTHIA as 9,400 photo-realistic frames at $1280 \times 960$ with 13 classes, which suggests that different papers may be referring to different benchmark conventions or internal preprocessing choices rather than to a single immutable specification (Yu et al., 2024).

The dataset is also used in sequence form. One point-cloud segmentation study describes the original SYNTHIA data as stereo RGBD images generated by 4 cameras located on the top of a moving car, and preprocesses 6 video sequences spanning 9 different weather environments into dynamic 3D point cloud sequences (Wang et al., 2020). A recurrent video-segmentation paper notes that SYNTHIA contains over 200,000 images with varying weather and seasons, but uses only a portion of the Highway sequence in summer conditions, again illustrating that most published results rely on task-specific subsets rather than on the entire corpus (Siam et al., 2016).

2. SYNTHIA in synthetic-to-real semantic segmentation

The dominant use of SYNTHIA is as a labeled synthetic source domain for unsupervised or semi-supervised transfer to real urban-scene datasets such as Cityscapes, Mapillary Vistas, and BDD. The typical setup is to train on labeled SYNTHIA images while treating the target-domain training split as unlabeled during adaptation and evaluating on the target validation split. The rationale is consistent across papers: SYNTHIA offers large-scale dense supervision “for free,” but models trained only on synthetic appearance statistics overfit and transfer poorly to real images because the source-to-target gap is substantial (Zhang et al., 2020, Brehm et al., 2021).

A recurrent theme is that SYNTHIA is not merely a convenient source of labels; it is a stress test for adaptation methods. One semantically consistent translation paper explicitly calls SYNTHIA a harder adaptation source than GTA5 because the domain gap is wider, with more varied viewpoints, fewer annotated images, and some missing classes (Brehm et al., 2021). Another self-supervised adaptation paper emphasizes that SYNTHIA differs from Cityscapes not only in appearance but also in viewpoint statistics and scene geometry, which makes pseudo-label quality especially fragile (Iqbal et al., 2022). This suggests that strong performance on SYNTHIA $\rightarrow$ Cityscapes is often interpreted as evidence of robustness to both texture and geometric shift.

The reported results are benchmark-specific and use different backbones, loss functions, and class protocols, so they are not directly interchangeable. Even so, the trajectory of published results is informative.

Benchmark setting	Reported result	Paper
SYNTHIA $\rightarrow$ Cityscapes, SPIGAN	34.7 mIoU at $320 \times 640$; SPIGAN-no-PI: 31.1	(Lee et al., 2018)
SYNTHIA $\rightarrow$ Cityscapes, FCN$_{adv}$+RPT$^3$+MS	51.7% mIoU@16, 59.5% mIoU@13	(Zhang et al., 2020)
SYNTHIA $\rightarrow$ Cityscapes, semantically consistent I2I + SSL	42.7% $\rightarrow$0, 49.8% $\rightarrow$1	(Brehm et al., 2021)
SYNTHIA $\rightarrow$2 Cityscapes, DRSL+	46.7 mIoU, 53.2 mIoU*	(Iqbal et al., 2022)
SYNTHIA $\rightarrow$3 Cityscapes, CICLD / ICCLD	67.2 mIoU	(Yu et al., 2024)

When a small amount of target supervision is allowed, SYNTHIA also serves as the source domain in semi-supervised domain adaptation. A 2024 SSDA study reports 64.5 / 73.9 with 50 labels and 67.2 / 75.7 with 100 labels on Synthia$\rightarrow$4Cityscapes, where the paired numbers correspond to 16 / 13 classes; the fully supervised references are 68.9 / 73.1 (Morales-Brotons et al., 2024). A plausible implication is that, in this benchmark family, a small labeled target subset can close much of the residual gap that classical UDA leaves open.

3. Privileged information, cross-task transfer, and auxiliary supervision

One of the most influential reinterpretations of SYNTHIA is as a source not only of labels but also of simulator-internal signals. In SPIGAN, SYNTHIA supplies synthetic triplets $\rightarrow$5, where $\rightarrow$6 is semantic ground truth and $\rightarrow$7 is depth from the simulator’s z-buffer. The framework jointly learns a generator $\rightarrow$8, discriminator $\rightarrow$9, task network $1280 \times 960$0, and privileged network $1280 \times 960$1, with the objective

$1280 \times 960$2

and weights $1280 \times 960$3, $1280 \times 960$4, $1280 \times 960$5, $1280 \times 960$6. The paper frames this as Learning Using Privileged Information (LUPI) for sim-to-real adaptation, with SYNTHIA acting simultaneously as appearance source, label source, and privileged-information source (Lee et al., 2018).

A related but distinct line of work augments semantic adaptation with additional derived signals. In an edge-aware entropy-based adversarial framework, SYNTHIA is again the labeled source domain, but now provides semantic labels $1280 \times 960$7, depth labels $1280 \times 960$8, and edge labels $1280 \times 960$9, where the edge supervision is extracted from the semantic ground-truth maps using Canny edge detection. The key design choice is to concatenate edge probabilities with entropy maps before domain discrimination; on the reported benchmarks this yields 46.1 mIoU for SYNTHIA $\rightarrow$0 Cityscapes 16 classes, 72.0 for the 7-class Cityscapes protocol, and 69.6 for the 7-class SYNTHIA $\rightarrow$1 Mapillary protocol (Hong et al., 2023). This directly addresses a common misconception that synthetic-to-real adaptation is only about aligning texture or feature distributions: boundary precision is treated here as a first-class adaptation signal.

SYNTHIA has also been used to learn relationships between tasks. In Across Tasks and Domains Transfer (AT/DT), it is the fully supervised source domain $\rightarrow$2 for semantic segmentation and monocular depth estimation. The framework trains task networks $\rightarrow$3 and then learns a transfer mapping $\rightarrow$4 on SYNTHIA before applying it to partially supervised target domains such as Cityscapes, KITTI, and Carla. In the reported SYNTHIA $\rightarrow$5 Cityscapes depth $\rightarrow$6 semantics setting, mIoU rises from 7.72 to 23.24 and accuracy from 28.49 to 64.03 (Ramirez et al., 2019). Here SYNTHIA is valuable not just because it is synthetic, but because it is fully supervised for multiple dense tasks.

4. Temporal, 3D, panoramic, and omnidirectional extensions

SYNTHIA has repeatedly been used to study whether scene understanding improves when the input departs from the single forward-facing frame. In video semantic segmentation, recurrent and multi-stream architectures exploit the temporal continuity of driving scenes. On SYNTHIA-CVPR'16, MSFCN-3 reports 94.38 mean IoU for the Highway scenario and 88.89 for the City scenario, compared with 85.42 and 73.88 for the single-frame FCN baseline; the same paper reports that temporal models especially improve thin or weak classes such as Sidewalk and Pole (Sistu et al., 2019). An earlier recurrent-FCN study on the Highway summer sequence reports FC-VGG: 0.755 mean class IoU and RFC-VGG: 0.812, a 5.7% absolute improvement, with especially large gains for Car and Pedestrian (Siam et al., 2016).

The dataset is equally important in dynamic 3D point-cloud segmentation. In ASTA3DCNN, SYNTHIA is converted from RGBD video into point-cloud sequences by constructing, for each frame, a $\rightarrow$7 cube centered on the moving car and then applying Farthest Point Sampling to retain 8,192 points per frame. The reported split uses 19,888 training frames, 815 validation frames, and 1,886 test frames. The segmentation network organizes each core point around 4 virtual anchors in a regular tetrahedron, and the anchor features are aggregated by

$\rightarrow$8

On SYNTHIA, the paper reports 84.26 mIoU for 2-frame input and 84.77 mIoU for 3-frame input, while also noting a persistent failure case: the Bicycle class remains at 0.00 IoU for all methods in the main table (Wang et al., 2020).

Panoramic and spherical derivatives extend SYNTHIA’s geometric coverage. SYNTHIA-PANO is built by stitching the left, forward, right, and backward views after cylindrical projection, yielding panoramic images of size (3340, 760) with 360-degree FoV. The paper reports that training with panoramic images is beneficial, and that 180-degree FoV is the best trade-off: performance improves from 90 to 180 degrees, but degrades at 360 degrees because of the extreme aspect ratio and receptive-field limitations of ICNet (Xu et al., 2019). The same paper also describes an internal inconsistency in the distortion discussion: one sentence says that when focal length is shorter than 625, $\rightarrow$9 performs better in some regimes, while a later conclusion says that when focal length is shorter than 625, $\rightarrow$0 always performs better. The stated overall conclusion is nevertheless that panoramic training improves anti-distortion robustness (Xu et al., 2019).

For omnidirectional segmentation, Omni-SYNTHIA samples SYNTHIA onto an icosahedron mesh with

$\rightarrow$1

At $\rightarrow$2 the proposed HexUNet reports 43.6 mIoU and 52.2 mAcc, outperforming UNet and UGSCNN; at $\rightarrow$3 it reaches 48.3 mIoU and 57.1 mAcc (Zhang et al., 2019). The paper emphasizes that the filters are orientation-aware rather than rotation invariant, because in upright road scenes sky tends to be “up” and road tends to be “down.” Even here, small objects remain difficult: pedestrian and sign are weak, and all methods fail for cyclist (Zhang et al., 2019).

5. Domain generalization and panoptic adaptation

Although SYNTHIA is most often discussed in UDA, it is also a strong testbed for domain generalization. A texture-randomization framework treats SYNTHIA-RANDCITYSCAPES as one of two synthetic source domains and reports 39.7 mIoU for SYNTHIA $\rightarrow$4 Cityscapes, 35.3 for SYNTHIA $\rightarrow$5 BDDS, and 36.4 for SYNTHIA $\rightarrow$6 Mapillary with ResNet-101. The method uses Global Texture Randomization (GTR), Local Texture Randomization (LTR), and a consistency term CGL, and the authors explicitly compare against domain-adaptation methods despite using no target-domain data during training (Peng et al., 2021). The central claim is that aggressive appearance diversification forces the model away from texture shortcuts and toward shape, layout, and context.

A second domain-generalization line exploits the fact that SYNTHIA contains multiple weather and environment shifts. ASH+ trains on Highway-Dawn, Highway-Fog, or Highway-Spring, and evaluates on unseen conditions across New York-like and Old European Town environments. The paper reports 36.30 for Highway (Dawn), 34.54 for Highway (Fog), and 41.39 for Highway (Spring), and identifies the best stylization-strength setting on SYNTHIA as $\rightarrow$7 (Tjio et al., 2023). This is presented as evidence that hallucination strength should be dataset- and pixel-dependent rather than globally fixed.

SYNTHIA has also become a benchmark for panoptic segmentation UDA. In EDAPS, it is the synthetic source domain for SYNTHIA $\rightarrow$8 Cityscapes and SYNTHIA $\rightarrow$9 Mapillary Vistas, with 9,400 synthetic images with panoptic labels. The supplementary description follows CVRN and uses 11 stuff classes and 8 thing classes. The proposed shared MiT-B5 transformer encoder with task-specific decoders reports mSQ 72.7, mRQ 53.6, mPQ 41.2 on SYNTHIA $320 \times 640$0 Cityscapes and mSQ 71.7, mRQ 46.1, mPQ 36.6 on SYNTHIA $320 \times 640$1 Mapillary Vistas (Saha et al., 2023). The gains over prior work are large, but the failure modes are also explicit: on Cityscapes, fence and bike remain very weak, and on Mapillary the fence class remains at 0 PQ (Saha et al., 2023). This reinforces a broader pattern visible across SYNTHIA-based research: thin structures, small objects, and rare classes remain difficult even with synthetic dense supervision.

6. Reuse of the name outside the driving benchmark

The label SYNTHIA / Synthia / SynthIA is no longer unique to the driving dataset. In QuPAINT, Synthia is a physics-based synthetic data generator for optical microscopy patches of two-dimensional quantum material flakes. It uses the Transfer Matrix Method (TMM), CIE 1931 color rendering, a white-balance-aware module, and a substrate-aware module, and the paper states that it generates 50,000 samples per material, for a total of 400,000 synthetic samples in the training set (Nguyen et al., 19 Feb 2026).

In generative design, SYNTHIA denotes a framework for novel concept design with affordance composition. Its ontology is reported as spanning 30 superordinates, 590 concepts, 1172 parts, and 686 affordances, and the abstract summarizes human-evaluation gains of 25.1% in novelty and 14.7% in functional coherence over prior methods (Ha et al., 25 Feb 2025). In persona-driven language modeling, SYNTHIA is a synthetic persona dataset with 30,000 backstories generated from 10,000 real BlueSky users across three temporal windows—10%, 50%, and 100% of each user’s latest two-year posting history (Rahimzadeh et al., 20 Jul 2025). In solar physics, SynthIA means Synthetic Inversion Approximation, a U-Net-based system that maps 25-channel SDO/HMI observation stacks to Hinode-like magnetic inversion products; the paper states that the model contains over 10 million trainable parameters and produces the SynodeP data product (Higgins et al., 2021).

Across the cited vision papers, however, SYNTHIA continues to denote a synthetic urban-scene source domain and its derivatives. The recurring technical significance of that benchmark is not simply that it is synthetic, but that it exposes several distinct research problems at once: dense supervision without manual labeling, simulator-internal privileged information, severe synthetic-to-real domain shift, multi-view and temporal structure, and controlled variations in weather, illumination, and geometry.