VehicleSeg10K: Benchmark for Vehicle Part Segmentation

• VehicleSeg10K is a benchmark dataset offering 11,665 real vehicle images and 97,471 pixel-level annotations across 13 semantic categories to support fine-grained segmentation.
• It is designed to tackle challenges in scale, viewpoint diversity, and annotation density, thereby enabling robust training for autonomous driving and smart transportation systems.
• The dataset underpins the SAV framework by integrating semantic segmentation with a vehicle part knowledge graph to improve structural understanding and model performance.

VehicleSeg10K is a large-scale benchmark dataset for vehicle part segmentation introduced in "Segment Any Vehicle: Semantic and Visual Context Driven SAM and A Benchmark" (Wang et al., 6 Aug 2025). It is designed for fine-grained, pixel-level parsing of vehicle exterior structure under realistic deployment conditions, and it serves both as the principal training and evaluation resource for the SAV framework and as the basis for a benchmark covering 18 segmentation algorithms. The dataset comprises 11,665 real vehicle images and 97,471 annotated part instances across 13 semantic categories, with explicit emphasis on scale, viewpoint diversity, nighttime imagery, and adverse weather conditions.

1. Origins and design objectives

VehicleSeg10K was constructed to address three gaps identified in prior vehicle-part datasets: insufficient scale and annotation density, limited diversity of real-world conditions, and inadequate support for fine-grained part semantics (Wang et al., 6 Aug 2025). The motivating comparison is with datasets such as UDA-Part and 3DRealCar-Part, which are characterized respectively as synthetic or constrained in viewpoints and environmental conditions. Within this framing, VehicleSeg10K is intended to support robust model training for autonomous driving, ADAS, intelligent traffic systems, damage assessment, automated parking, and other vehicle-centric applications requiring part-level understanding.

The dataset is explicitly organized around semantic segmentation of individual vehicle parts rather than whole-vehicle masks. This focus reflects the claim that higher-level vehicle perception tasks require discrimination among doors, windows, wheels, license plates, and body regions. The 13-category ontology therefore functions not only as an annotation schema but also as a semantic prior for structured modeling.

The paper further positions VehicleSeg10K as a foundation benchmark for fine-grained vehicle part segmentation, for context- and knowledge-graph–aware models such as SAV, and for systematic comparison under diverse realistic conditions. The statement that it is "the first large-scale vehicle part segmentation dataset that simultaneously addresses scale, diversity, and real-world deployment challenges" should be read as the paper’s own characterization of its contribution.

2. Scale, splits, and semantic schema

VehicleSeg10K contains 11,665 real vehicle images and 97,471 part annotations across 13 categories (Wang et al., 6 Aug 2025). The annotation target is a per-pixel semantic label over a fixed part vocabulary:

1. Foreground
2. Wheel
3. Plate
4. Front window
5. Back window
6. Left front window
7. Left front door
8. Left back window
9. Left back door 10. Right front window
10. Right front door
11. Right back window
12. Right back door

The "Foreground" category denotes the vehicle body excluding the listed specific parts. These 13 classes are the semantic categories used for segmentation and simultaneously the node set of the vehicle part knowledge graph in SAV.

The paper presents two different train/test split descriptions. In the detailed dataset section, it states that "The entire dataset is divided into 8,596 training images and 2,150 testing images to support model training and evaluation processes." Elsewhere, it mentions "training and testing subsets, which contain 8,596 and 3,069 images, respectively." The detailed section is the more explicit and self-consistent statement. Since 8,596 plus 2,150 does not equal 11,665, this suggests that 919 images are not used in the public benchmark split, although no explicit validation subset is named.

Image resolutions range from $229 \times 202$ to $1411 \times 1302$ pixels. During SAV training, images are resized to $640 \times 480$ for efficiency, but the dataset itself preserves heterogeneous native resolutions. This combination of native variability and standardized training preprocessing implies a benchmark regime in which spatial scale variation remains part of the data distribution even if model inputs are normalized during optimization.

3. Data collection and annotation protocol

VehicleSeg10K is composed of real images collected by web crawling under data privacy protection measures (Wang et al., 6 Aug 2025). The text does not enumerate the specific websites or collection jurisdictions, but it indicates deliberate curation for six vehicle types, multiple viewpoints, day and night scenes, adverse weather, and broad resolution variation.

The six major vehicle types are:

• Sedan
• SUV
• MPV
• Van
• Sports Car
• Pickup

Annotations are pixel-level and assign each part category a unique label ID. The paper describes the masks as "high-quality pixel-level annotations" and "pixel-level precise segmentation masks," with storage in standard formats compatible with deep learning frameworks. The initial labeling is performed by student annotators under expert supervision. Quality control is multi-stage: specialized reviewers assess annotation quality after the initial pass, and final expert validation checks the refined masks. The text emphasizes that the annotations are "carefully annotated and refined," indicating an explicit effort to maintain inter-image consistency for semantically adjacent or visually similar parts.

Although the paper does not provide per-class annotation difficulty statistics, it acknowledges challenging cases implicitly through the training design used on the dataset. Distinguishing adjacent parts with similar appearance, such as windows and doors, and handling small imbalanced regions such as license plates are treated as nontrivial aspects of the benchmark. This suggests that VehicleSeg10K should not be interpreted merely as a large mask repository, but as a dataset constructed around fine-grained boundary precision and semantically dense part interactions.

4. Diversity, distributions, and relation to earlier datasets

A central property of VehicleSeg10K is its explicit coverage of realistic variation in viewpoint, illumination, weather, vehicle type, and color (Wang et al., 6 Aug 2025). The dataset includes multi-angle views, including both horizontal and elevated views with "comprehensive vertical angle variations." It also contains daytime and nighttime scenes, outdoor real-world backgrounds, and adverse weather conditions including rain, fog, snow, and dust storms.

The paper’s statistical analysis highlights several distributional characteristics. Wheels are the most common annotated component, followed by the general "Foreground" category and window/door categories. Sedans and SUVs constitute the majority of vehicle types. A color distribution plot is described as reflecting realistic distributions of vehicle body colors, and a vehicle center point heatmap shows typical positioning of vehicles within images, providing an explicit view of composition bias.

Relative to UDA-Part and 3DRealCar-Part, VehicleSeg10K is presented as both complementary and broader in deployment-oriented diversity. UDA-Part has 31 part IDs, 4 vehicle types, and 3,862 images, but is synthetic and uses fixed viewpoints. 3DRealCar-Part has 13 part IDs, 6 vehicle types, and 10,219 images at $1920 \times 1440$ resolution, but is described as "Horizontal only" in viewpoint and day-only in scenario coverage. VehicleSeg10K matches 3DRealCar-Part in part-level granularity, exceeds it in image count, and extends coverage to nighttime scenes and adverse weather. In that sense, the benchmark is configured not primarily around the maximum number of labels, but around realism, heterogeneity, and robustness-relevant nuisance factors.

5. Coupling to the SAV knowledge graph

VehicleSeg10K’s label ontology is directly embedded in the SAV framework through a Vehicle Part Knowledge Graph formalized as

$G = (V, E, W),$

where $V = \{v_1, v_2, \ldots, v_n\}$ are vehicle part nodes, $n = 13$ for VehicleSeg10K, $E$ denotes edges representing physical adjacency, and $W \in \mathbb{R}^{n \times n}$ is the adjacency matrix with learned edge weights (Wang et al., 6 Aug 2025).

Each part class $i$ is initialized from a CLIP text embedding

$1411 \times 1302$0

obtained by encoding its textual label. Edges are defined according to the physical or structural topology of the vehicle. The paper gives the example that the left front door connects to the left front window, left back door, and foreground, while maintaining no connection to right-side components because of the vehicle’s bilateral symmetry. This construction makes the graph closer to a topology graph than to a fully connected semantic co-occurrence graph.

The edge weights are derived from co-occurrence statistics in the training portion of VehicleSeg10K: $1411 \times 1302$1 where $1411 \times 1302$2 is the number of training images in which both parts appear and $1411 \times 1302$3 is the total number of training images. The dataset therefore contributes not only supervised masks but also empirical structural statistics that shape the model prior.

Graph encoding is performed with a 4-layer GATv2: $1411 \times 1302$4 The attention mechanism is defined as

$1411 \times 1302$5

$1411 \times 1302$6

This coupling means that VehicleSeg10K is structurally integral to SAV rather than merely a benchmark on which SAV is tested. A plausible implication is that part definitions, co-occurrence statistics, and topology are jointly operationalized: the dataset’s ontology defines the nodes, the training split defines empirical co-occurrence weights, and the model architecture consumes both as priors for segmentation.

6. Benchmark task, metrics, and reported results

The primary task defined on VehicleSeg10K is semantic segmentation of vehicle images into the 13 part classes (Wang et al., 6 Aug 2025). In SAV, the decoder outputs 13 class-specific masks in a single forward pass: $1411 \times 1302$7

Evaluation is reported using mean Intersection over Union and mean Accuracy. For $1411 \times 1302$8 classes, the classwise IoU is

$1411 \times 1302$9

and

$640 \times 480$0

Per-class accuracy is

$640 \times 480$1

and

$640 \times 480$2

The benchmark retrains and evaluates 18 baseline methods in addition to SAV. The reported VehicleSeg10K results are:

Method	mIoU (%)	mAcc (%)
DeepLabV3	66.18	77.56
DeepLabV3+	66.23	77.66
DDRNet	64.12	75.90
PoolFormer	72.08	82.51
ConvNeXt	72.58	82.50
SegNeXt	64.26	76.26
MaskDINO	73.50	83.68
GSS-gss	57.89	70.72
PEM	61.53	72.56
FADC	71.90	81.84
HieraHyp	64.13	74.96
CGRSeg	70.01	80.67
Spike2Former	61.19	72.55
gcnet	73.40	83.11
ProPETL	74.48	84.14
SegMAN	62.69	73.97
CWSAM	76.90	82.38
WPS-SAM	74.27	83.04
SAV	81.23	87.08

Within this benchmark, CWSAM is the strongest non-SAV method in mIoU at 76.90, whereas ProPETL has the highest non-SAV mAcc at 84.14. SAV reaches 81.23 mIoU and 87.08 mAcc, a gain of +4.33 mIoU over CWSAM. The paper interprets this as evidence that combining semantic structure and visual context is particularly beneficial for vehicle part segmentation.

The training strategy used for these fine-grained labels includes a combined loss of binary cross-entropy mask loss, Dice loss, and a weighted classification loss, along with uncertainty-guided point sampling focused on regions with high prediction uncertainty, particularly near boundaries. In context, these design choices reflect the dataset’s emphasis on adjacent-part discrimination and small-region segmentation.

7. Intended uses, limitations, and availability

VehicleSeg10K is intended for fine-grained vehicle part segmentation, context- and knowledge-graph–enhanced modeling, benchmarking of segmentation methods, and downstream applications in autonomous driving and intelligent transportation, including damage assessment, automated parking assistance, insurance or maintenance analysis, and high-level scene understanding (Wang et al., 6 Aug 2025). Because the benchmark includes both conventional CNN-based segmentation models and recent transformer or SAM-related architectures, it also functions as a comparative testbed for architectural trends in structured part parsing.

Several limitations are either explicit or strongly implied. The vehicle center point heatmap indicates a composition bias toward typical image positions. The reported color and type distributions reflect authentic real-world frequencies, which may underrepresent rarer vehicle types or colors. The ontology covers 13 exterior part classes but does not separately annotate mirrors, lights, or interior components. For applications requiring more granular exterior decomposition or cabin-level understanding, the schema would need extension. The paper does not explicitly analyze geographic provenance; a plausible implication of web-crawled collection is that regional bias may exist in vehicle models, license plates, and traffic environments.

There is also a textual ambiguity in the split description, as noted earlier, and the paper does not identify a formal validation subset. This is not a criticism of the benchmark definition itself, but it is a relevant detail for reproducibility and leaderboard interpretation.

The paper states that both the dataset and source code will be released at: $640 \times 480$3 and describes the materials as open-sourced. No specific license is given in the text. Accordingly, the repository is the intended source for images, pixel-level masks, pre-trained SAV models, baseline training and evaluation code, and any accompanying data-processing scripts.