Organizations: Beijing Institute of Technology · Tsinghua University · China Agricultural University · Beijing Jiaotong University · The Hong Kong Polytechnic University · Li Auto
Multi-task visual anomaly detection is critical for car-related manufacturing quality assessment. However, existing methods remain task-specific, hindered by the absence of a unified benchmark for multi-task evaluation. To fill in this gap, We present the CAD Dataset, a large-scale and comprehensive benchmark designed for car-related multi-task visual anomaly detection. The dataset contains over 100 images crossing 7 vehicle domains and 3 tasks, providing models a comprehensive view for car-related anomaly detection. It is the first car-related anomaly dataset specialized for multi-task learning(MTL), while combining synthesis data augmentation for few-shot anomaly images. We implement a multi-task baseline and conduct extensive empirical studies. Results show MTL promotes task interaction and knowledge transfer, while also exposing challenging conflicts between tasks. The CAD dataset serves as a standardized platform to drive future advances in car-related multi-task visual anomaly detection.
Figures & tables
Figure 1: Overview of the CAD 100K Dataset. The CAD 100K dataset contains over 100 K images crossing 7 domains and 3 tasks, providing a comprehensive view for car-related anomaly detection. It features a multi-task architecture and synthesis data augmentation methods, solves cross-domain problems, supports both supervised and unsupervised learning paradigms, and is compatible with few-shot learning frameworks.
Figure 2: Tasks and structure of the CAD 100K Dataset. CAD 100K adopts a hierarchical domain–system–part anomaly structure to ensure semantic consistency and extensibility across different tasks.
Figure 3: The data composition across domains, tasks, and sources in CAD 100K. All images are in RGB format with various resolutions from 100 × 84 to 4096 × 3072, covering diverse lighting and viewpoints.
Figure 4: Structure of the proposed Baseline.
Backbone Type
Backbone Size
Method
Cls. Acc (%)
Det. mAP (%)
Seg. mIoU (%)
DINOv3 (ViT)
Smallplus
Single
98.40
59.20
72.80
Base
98.40
58.90
70.93
Adaptive
98.40
57.80
71.88
Base
Single
100.00
59.40
72.42
Base
98.80
58.20
71.95
Adaptive
99.10
58.90
72.25
Table 1: Comprehensive performance comparison on public car-related benchmarks across different backbone architectures, scales, and training paradigms. Results show classification accuracy (Cls. Acc), detection mAP (Det. mAP), and segmentation mIoU (Seg. mIoU). Bold indicates the best performance within each backbone size and architecture combination.
Backbone
Method
Cls. Acc (%)
Det. mAP (%)
Seg. mIoU (%)
DINOv3-Splus
Base
92.98
55.20
56.84
Adaptive
89.47
59.60
57.19
ConvNeXt-S
Base
92.10
59.10
52.20
Adaptive
90.10
60.10
52.77
Table 2: Performance comparison on three types of tasks in the CAD 100K dataset configurations: ConvNeXt-S with base and adaptive MTL, and DINOv3-Splus with base and adaptive MTL. Results demonstrate the effectiveness of our adaptive strategy across different backbone architectures.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B1 : The Exterior Domain Subset Supporting both Detection and Segmentation in the CAD 100K Dataset. This subset contains various exterior anomalies with both bounding box and pixel-level mask annotations.
Figure B2 : The Interior and General Domain Detection Subset in the CAD 100K Dataset. Examples include interior component defects and general cross-system anomalies with detection annotations.
Figure B3 : The Chassis Domain Classification Subset in the CAD 100K Dataset. Covering suspension, brake, and tire systems with various anomaly types.
Figure B4 : The Engine Domain Classification Subset in the CAD 100K Dataset. Featuring engine components, fluid systems, and mechanical parts.
Figure B5 : The Electrical Domain Classification Subset in the CAD 100K Dataset. Including battery systems, wiring harnesses, and electronic components.
Figure B6 : The EV-specific Domain Classification Subset in the CAD 100K Dataset. Specialized components for electric vehicles including high-voltage systems and charging infrastructure.
Model
Interior
General
Exterior
AP
AP 50
AP 75
AR
AP
AP 50
AP 75
AR
AP
AP 50
AP 75
AR
DINOv3-Splus
48.0
79.9
46.7
61.0
62.4
91.2
65.88
68.9
45.1
66.9
45.2
54.5
ConvNeXt-S
46.9
76.1
46.6
61.8
64.8
92.8
72.3
72.4
51.0
72.0
54.4
57.6
Appendix
Table B1 : Detection Performance Comparison Across Vehicle Domains . Models are evaluated on three vehicle domains (Interior, General, Exterior) using standard COCO metrics including AP at different IoU thresholds and AR.
Model
Chassis
Engine
Electric
EV-specific
Acc@1
Acc@5
Acc@1
Acc@5
Acc@1
Acc@5
Acc@1
Acc@5
DINOv3-Splus
91.2
100.0
90.1
100.0
93.5
99.1
97.9
100.0
ConvNeXt-S
90.4
100.0
93.4
100.0
93.5
100.0
100.0
100.0
Appendix
Table B2 : Classification Performance Across Vehicle Component Domains . Models are evaluated on four vehicle component domains using top-1 and top-5 accuracy metrics.
Model
Car Exterior Anomalies
Car Parts Segmentation
mIoU
mFscore
mPrec
mRecall
mIoU
mFscore
mPrec
mRecall
DINOv3-Splus
54.9
67.0
85.7
58.1
72.0
81.7
81.9
81.9
ConvNeXt-S
52.4
65.1
90.1
54.4
72.9
82.4
82.7
82.6
Appendix
Table B3 : Segmentation Performance Comparison . Models are evaluated on both car exterior anomalies segmentation and car parts segmentation tasks using mIoU, mFscore, and mPrecision metrics.
Model
Method
Cls. Acc (%)
Det. mAP (%)
Seg. mIoU (%)
DINOv3-Splus
Single
98.40
59.20
72.80
Base-MTL
98.40
58.90
70.93
Adaptive-MTL
98.40
57.80
71.88
ConvNeXt-S
Single
98.80
58.80
72.80
Base-MTL
98.80
58.60
71.12
Adaptive-MTL
98.80
59.10
71.82
Appendix
Table B4 : Multi-task Learning Performance Comparison on Public Benchmarks . Models are evaluated across classification, detection, and segmentation tasks using single-task learning (STL), base multi-task learning (Base-MTL), and adaptive multi-task learning (Adaptive-MTL) approaches.
Department of Computer Science and Engineering, Southern University of Science and Technology, Shenzhen, China · Macau University of Science and Technology, Macau, China · Tencent YouTu Lab, Shenzhen and Shanghai, China +3