Enterprise AI Analysis

2. Model Architecture

Contemporary object detection architectures exhibit a fundamental trade-off between semantic representation and spatial localization accuracy, where convolutional neural network (CNN)-based detection frameworks like the YOLO family demonstrate superior spatial localization through supervised learning but face three critical limitations. Supervised detectors learn dataset-specific feature representations optimized for training distributions, performing poorly on out-of-distribution scenarios such as unconventional illumination, partial occlusions, or novel viewpoints—particularly problematic for domain-specific applications like construction safety monitoring and infrastructure inspection where the COCO dataset's 118,000 annotated images sets an impractical benchmark for specialized tasks typically possessing fewer than 200 samples. Under such data-scarce conditions, supervised YOLO architectures exhibit substantial performance degradation through overfitting and insufficient feature diversity. Concurrent with supervised detection research, self-supervised learning has emerged as a paradigmatic shift, with DINOv3—a 7-billion parameter Vision Transformer trained on 1.7 billion unlabeled images—demonstrating that self-supervised models can outperform weakly-supervised counterparts by learning rich, transferable visual semantics without classification taxonomies. By leveraging DINOv3's frozen backbone features, detection systems can achieve robust performance with minimal annotated images—a regime where traditional supervised architectures fail—democratizing advanced object detection for specialized applications where large-scale annotation is infeasible.

2.1 Architectural Modification Rationale

The complementary characteristics of supervised detection architectures and self-supervised visual representation models motivate the central research hypothesis: the supervised localization optimization of YOLO architectures can be synergistically integrated with the self-supervised semantic representations of DINOv3 to construct a hybrid detector exhibiting both spatial precision and semantic robustness under data-scarce training conditions. Previous integration methodologies exhibit fundamental limitations. Complete backbone substitution eliminates YOLO's multi-scale feature pyramid hierarchy, substantially increases computational complexity, and degrades small object detection performance. Single-point feature injection constrains semantic enhancement to a singular hierarchical level, precluding multi-scale semantic enrichment throughout the detection pipeline.

The proposed architectural modification implements hierarchical semantic feature injection at two strategically selected positions. Input-level integration at the initial feature extraction stage (P0) transforms raw pixel representations into semantically-grounded feature spaces prior to spatial downsampling, ensuring all subsequent processing benefits from enhanced semantic priors. Mid-level integration at the intermediate feature pyramid level (P3) provides semantic enhancement at the architectural position representing optimal balance between semantic abstraction and spatial resolution, directly augmenting feature quality for detection head processing. This dual-injection strategy instantiates the Hierarchical Semantic Specialization Principle: architectural levels processing information at distinct abstraction granularities require qualitatively different forms of semantic enhancement to maximize detection performance under limited training data conditions.

2.2. Integration Location Rationale

Figure 1 illustrates the proposed YOLOv12-L architecture with dual-level DINOv3 integration at P0 and P3 positions. The architecture processes a 640×640 input image through a hierarchical pipeline comprising input preprocessing (Layer 0), early backbone stages (Layers 1-5), DINOv3-enhanced mid-backbone (Layer 6), deep backbone stages (Layers 7-10), feature pyramid network with top-down and bottom-up pathways (Layers 11-22), and multi-scale detection heads operating at P3/8, P4/16, and P5/32 resolutions (Layer 23). Two DINOv3-ViT-B/16 modules (86M parameters each, frozen) are strategically positioned: the DINO3Preprocessor at P0 transforms raw RGB inputs into semantically-enriched 3-channel representations, while the DINO3Backbone at P3 enhances 512-channel feature maps at 80×80 spatial resolution through global self-attention mechanisms with gated fusion and residual connections. The complete architecture contains 220M total parameters with 47M trainable parameters (21%), operating at 135 GFLOPs for 640×640 inputs. This dual-injection strategy is motivated by distinct functional requirements at different abstraction levels within the detection pipeline, addressing complementary aspects of semantic representation learning under data-scarce conditions.

2.2.1. P0 (Input Preprocessing Layer) Integration

The input preprocessing stage (Layer 0 in Figure 1) represents the sole architectural location where feature enhancement propagates through all subsequent network layers. This position enables transformation of raw pixel representations into semantically-grounded features prior to any task-specific processing, establishing semantic priors that influence the entire network hierarchy. P0 integration addresses the fundamental challenge that supervised detectors trained on limited data learn features from raw pixel intensities without leveraging prior visual knowledge, resulting in inefficient feature learning and poor generalization. The standard detection pipeline produces predictions according to Equation 1:

P(x) = f_head (f_backbone (f_input(x))) (1)

where x ∈ R^H×W×3 represents the raw image. Conventional architectures implement f_input as an identity transformation. The proposed modification replaces this identity mapping with π(D(x)), where D: R^H×W×3 → R^H×W×768 represents DINOv3 feature extraction and π: R⁷⁶⁸ → R³ denotes a learned projection reducing dimensionality to maintain architectural compatibility. As shown in Figure 10, the DINO3Preprocessor extracts 768-dimensional embeddings, upsamples spatial dimensions, and projects to 3 channels, outputting semantically-enhanced 640×640 representations that maintain dimensional compatibility with subsequent layers.

DINOv3's pretraining on 1.7 billion diverse images enables encoding of robust low-level visual primitives—textures, edges, color distributions—within semantically meaningful contexts rather than as isolated pattern responses. This semantic grounding of low-level features is critical for small-dataset scenarios where supervised training cannot adequately learn the relationship between primitive visual patterns and object semantics. By transforming the input space, P0 integration ensures that the first convolutional layer (Layer 1 in Figure 1: Conv 3→64, stride=2) receives semantically-coherent inputs rather than raw pixel intensities, accelerating convergence and improving generalization. Furthermore, self-supervised features exhibit inherent robustness to photometric transformations due to augmentation-invariant training objectives, with this robustness propagating throughout the network when injected at the input level. The dimensional projection from 768 to 3 channels maintains compatibility with YOLO's initial convolutional layer without downstream architectural modifications while forcing the projection layer to learn compact information-rich representations. Expected benefits include enhanced generalization to out-of-distribution data, improved robustness to appearance variations, and accelerated training convergence through semantically-meaningful initialization.

2.2.2. P3 (Mid-Backbone) Integration

While P0 integration provides semantic enhancement at the input level, it operates on raw visual primitives and undergoes substantial transformation through subsequent convolutional layers (Layers 1-5 in Figure 1). P3 integration (Layer 6 in Figure 1) addresses a complementary requirement: direct enhancement of mid-level feature abstractions at the architectural position where spatial resolution (80×80 grid) and semantic abstraction achieve optimal balance for object detection tasks. This position corresponds to the P3/8 feature pyramid level, which serves as a critical intermediate representation for small-to-medium object detection. The P3 level represents the first feature pyramid stage where sufficient spatial downsampling has occurred to enable global contextual reasoning (80×80 = 6,400 tokens for transformer processing) while maintaining adequate spatial resolution for precise object localization. At this hierarchical position, features have undergone initial abstraction through P1-P2 processing (Layers 1-3: 64→128→256 channels) but have not yet been subjected to aggressive spatial compression characteristic of deeper layers (P4: 40×40 in Layer 10, P5: 20×20). This intermediate abstraction level is precisely where semantic enhancement provides maximal benefit: features are sufficiently abstract to benefit from high-level semantic knowledge, yet retain sufficient spatial granularity to preserve localization accuracy. As illustrated in Figure 1, the DINO3Backbone module receives 512-channel features at 80×80 resolution from Layer 5, projects to 768 dimensions for DINOv3 transformer processing, applies global self-attention across 6,400 tokens, implements gated fusion with residual connections, and projects back to 512 channels, producing DINO-enhanced P3 features that maintain dimensional consistency with the original YOLO architecture.

The dual-injection strategy instantiates complementary enhancement mechanisms visible in the architectural flow of Figure 1. P0 enhancement operates on low-level visual primitives (edges, textures, colors), transforming the fundamental visual vocabulary from which all subsequent features are constructed, thereby establishing semantic groundwork that benefits feature learning at all hierarchical levels through Layers 1-23. P3 enhancement operates on mid-level feature abstractions after initial convolutional processing, directly enriching the feature representations consumed by the detection head and providing task-relevant semantic augmentation at the scale most critical for object detection performance. The architectural position of P3 ensures that semantic enhancement occurs immediately before the feature pyramid network pathways (Layers 11-22), enabling semantically-enriched features to propagate through both top-down (Layers 11-16) and bottom-up (Layers 17-22) pathways to the multi-scale detection heads at P3/8 (80×80, small objects 0-32² pixels), P4/16 (40×40, medium objects 32²-96² pixels), and P5/32 (20×20, large objects 96²+ pixels). In small-dataset scenarios where supervised learning cannot adequately capture semantic relationships between feature patterns and object categories, P3 integration provides direct semantic supervision through frozen pretrained features at the precise hierarchical level where detection decisions are formulated. This dual-level integration strategy embodies the Hierarchical Semantic Specialization Principle: architectural positions processing information at distinct abstraction granularities require qualitatively different forms of semantic enhancement to maximize detection performance under limited training data conditions.

3. Data Characteristics and Selection Rationale

The dataset selection strategy is predicated on evaluating DINO-YOLO's performance across progressively increasing data scales, from severely data-constrained to data-abundant regimes. This systematic approach enables quantification of the model's data efficiency and establishes its applicability across the spectrum of civil engineering deployment scenarios. The four datasets are strategically ordered by training sample size to characterize model behavior under varying data availability conditions, providing a logarithmic progression that spans three orders of magnitude from 648 to 118,000 training samples as shown in Table 1. This experimental design facilitates systematic investigation of the hypothesis that self-supervised pre-training provides disproportionate performance advantages in data-limited scenarios while maintaining competitive efficiency in data-rich environments.

The Tunnel Segment Crack dataset, comprising 648 training images, 50 validation images, and 40 test images, represents the most severe data scarcity scenario commonly encountered in specialized infrastructure inspection tasks. This dataset addresses quality control inspection of precast concrete tunnel segments during manufacturing or prior to installation, where early crack detection is critical for preventing defective segments from entering the tunnel assembly process. The extreme data limitation reflects practical constraints in precast concrete manufacturing facilities where defect samples are inherently rare (<2% rejection rates) and systematic image acquisition requires integration with production workflows. The single-class detection task (crack) presents extraordinary complexity due to extreme visual heterogeneity across the dataset: segments are captured in diverse storage configurations at highly variable distances (0.3-10m) from multiple viewing angles under inconsistent illumination conditions. The background environments present substantial visual clutter including steel reinforcement cages, storage racks, industrial equipment, and stacked adjacent segments. The concrete segment surfaces exhibit intricate visual complexity with casting artifacts, formwork seam lines, embedded bolt holes, surface texture variations, efflorescence deposits, and construction joints—all potentially confused with actual cracks. The crack instances demonstrate substantial morphological diversity ranging from hairline micro-cracks (<0.3mm width) barely visible at typical inspection distances to branching crack patterns with multiple propagation directions, creating optimization difficulties from severe class imbalance (<0.5% positive detection windows) combined with asymmetric loss requirements.

The Construction PPE dataset, containing 1,132 training images, 143 validation images, and 141 test images, exemplifies the limited data regime characteristic of domain-specific safety monitoring applications in construction environments. This intermediate scale reflects practical constraints where manual annotation by certified safety inspectors is feasible but resource-intensive, with each image requiring 5-10 minutes of expert labeling time to accurately identify 11 distinct equipment classes. The classification taxonomy encompasses both positive detection (presence of helmets, safety vests, gloves, boots, goggles, and other protective equipment) and negative detection (missing or improperly worn equipment), introducing asymmetric loss considerations for safety-critical applications where false negatives (missed violations) carry significantly higher consequences than false positives. The dataset captures diverse construction site conditions including varying lighting (outdoor daylight, indoor artificial lighting, shadowed areas), occlusion patterns (partial equipment visibility due to worker positioning or environmental obstacles), scale variations (workers at different distances from camera), and background complexity (cluttered construction environments with machinery, materials, and structures). These challenging conditions reflect operational deployment requirements for automated safety compliance monitoring systems where detection accuracy must exceed 95% to gain acceptance from safety managers and regulatory authorities.

The KITTI dataset, with 5,233 training images, 1,945 validation images, and 746 test images, constitutes the moderate data scale representing typical research datasets in civil engineering autonomous vehicle applications. This dataset was collected using a vehicle-mounted sensor suite comprising stereo cameras, Velodyne HDL-64E rotating 3D laser scanner, and GPS/IMU navigation system, providing synchronized multimodal data for comprehensive scene understanding. The 8-class urban scene taxonomy includes cars, vans, trucks, pedestrians, cyclists, trams, and miscellaneous objects, with 3D bounding box annotations specifying object location, orientation, and occlusion level. The dataset encompasses diverse driving scenarios including urban streets, residential areas, highways, and campus environments in Karlsruhe, Germany, captured under varying weather conditions (sunny, overcast, light rain) and illumination (daytime only, avoiding low-light complications). For civil engineering applications, KITTI serves as a benchmark for autonomous construction equipment navigation, mobile crane operation monitoring, and worksite traffic management systems where real-time detection (>10 FPS) is essential for collision avoidance and path planning. The moderate dataset size represents a transitional regime where conventional CNN architectures begin to achieve acceptable performance (>70% mAP) but still benefit substantially from architectural improvements and transfer learning strategies.

The COCO dataset, comprising 118,000 training images and 5,000 validation images from the 2017 release, establishes the data-abundant baseline for comparison with state-of-the-art general object detection methodologies. This large-scale dataset contains 80 object categories spanning everyday objects (person, vehicle, animal), furniture, appliances, food items, sports equipment, and household items, with instance-level segmentation masks and comprehensive metadata including occlusion flags, truncation indicators, and crowd annotations. The images are sourced from Flickr with diverse photographic conditions, aspect ratios, resolutions, and compositional complexity, providing extensive coverage of visual appearance variations for each object category. Unlike domain-specific civil engineering datasets, COCO includes minimal examples of construction equipment, safety gear, or infrastructure elements, making it a pure evaluation of transfer learning effectiveness from general visual recognition to specialized civil engineering tasks. At this data scale (>100K images), performance typically approaches asymptotic limits for supervised learning, where additional training data provides diminishing marginal returns and model architecture becomes the primary determinant of detection accuracy. The inclusion of COCO validates that DINO-YOLO's architectural enhancements do not compromise performance in data-rich scenarios while demonstrating that self-supervised pre-training benefits persist even when abundant annotated data is available, albeit with reduced relative improvement compared to data-constrained regimes.

This progressive evaluation framework provides three critical assessment intervals with distinct learning dynamics. The extreme scarcity regime (<1K images) tests whether self-supervised features can substitute for extensive annotated data, potentially reducing annotation requirements by an order of magnitude for viable model deployment in quality control applications where defect samples are inherently rare. The practical limitation regime (1-10K images) represents the most common operational constraint in civil engineering applications, where annotation budgets support thousands rather than tens of thousands of labeled examples. The relative abundance regime (>10K images) establishes performance ceilings and validates that efficiency gains in limited-data scenarios do not incur performance penalties when data becomes available. The logarithmic spacing between dataset sizes enables identification of critical transition points where architectural advantages manifest most strongly, informing cost-benefit analyses for annotation investment versus algorithmic development in resource-constrained civil engineering deployment contexts. This structured dataset selection enables systematic validation of the hypothesis that DINOv3 self-supervised vision transformers provide disproportionate advantages in data-constrained civil engineering applications while maintaining competitive performance and computational efficiency in data-rich scenarios, ultimately establishing DINO-YOLO as a practical solution for the full spectrum of civil engineering computer vision deployment requirements.

4. Experiment

The experimental results demonstrate substantial performance improvements of DINO-YOLO variants over baseline YOLOv12 architectures across all three datasets, with improvement magnitude correlating strongly with dataset size and task complexity as shown in Table 2 and Fig. 3. The performance gains are most pronounced on the KITTI dataset, where L-ViT-B-Dual achieves 72.06% mAP@0.5 compared to the best baseline YOLOv12-S at 41.54%, representing a remarkable 73.5% relative improvement. This dramatic improvement validates that self-supervised DINOv3 features provide maximal benefit in the moderate data regime (5,233 training images), where conventional supervised learning has insufficient data for robust feature learning but sufficient examples to effectively fine-tune pre-trained representations. However, this performance improvement comes with increased computational cost: L-ViT-B-Dual requires 33.25ms inference time compared to YOLOv12-S's 8.37ms, representing a 4× latency increase that remains acceptable for most civil engineering applications requiring real-time processing at 15-30 FPS. On the NVIDIA RTX 5090, this translates to approximately 30 FPS (1000ms/33.25ms) for the L-ViT-B-Dual configuration, maintaining real-time capability suitable for field deployment in construction site monitoring and autonomous vehicle navigation systems.

The Construction PPE dataset exhibits moderate performance improvements, with M-ViT-L-Dual achieving 55.77% mAP@0.5 compared to baseline YOLOv12-M at 49.04%, representing a 13.7% relative improvement. This intermediate improvement aligns with the dataset's limited data regime (1,132 training images), where self-supervised pre-training provides significant but not transformative benefits. The more modest gains compared to KITTI reflect the increased task complexity of 11-class detection with challenging occlusion patterns and the requirement to detect both equipment presence and absence. The M-ViT-L-Dual variant achieves this performance with 28.09ms inference time—a 3.3× increase over YOLOv12-M's 8.45ms but maintaining sufficient throughput for multi-camera construction site monitoring where 10-15 camera streams require processing.

The Tunnel Segment Crack dataset shows meaningful improvements across DINO-YOLO variants despite the extreme scarcity regime. For YOLOv12-L baseline (48.31%), the L-ViT-B-Dual configuration achieves 54.28% mAP@0.5, representing a 12.4% relative improvement. For YOLOv12-M baseline (53.42%), the M-ViT-B-Dual variant achieves _56.18% mAP@0.5, representing a 5.2% relative improvement. The highest absolute performance is obtained with M-ViT-L-Dual at 55.48% mAP@0.5, showing a 3.9% improvement over the YOLOv12-M baseline. These moderate gains in the extreme scarcity regime (648 training images) can be attributed to the extreme visual complexity of tunnel segment inspection environments, severe class imbalance (<0.5% positive detection windows), and the specialized nature of crack detection requiring fine-grained discrimination between hairline defects and confounding concrete surface features such as formwork lines and aggregate boundaries. Notably, the baseline YOLOv12-M already achieves relatively strong performance (53.42%), suggesting that conventional CNN architectures capture relevant features for this task, though DINO-YOLO variants consistently improve upon these baselines. The M-ViT-B-Dual variant achieves 21.18ms inference time—a 2.5× increase over baseline YOLOv12-M (8.45ms) but enabling real-time inspection workflows at 47 FPS suitable for automated manufacturing quality control systems.Comparing DINO-YOLO architectural configurations reveals important insights about optimal integration strategies. The Dual Integration strategy consistently achieves strong performance across all datasets, with L-ViT-B-Dual excelling on KITTI (72.06%), M-ViT-L-Dual leading on PPE (55.77%), and M-ViT-B-Dual performing best on Tunnel (56.18%). The Triple Integration variant (S-ViT-B-Triple) shows weaker performance across all datasets (43.96% Tunnel, 68.16% KITTI, 47.27% PPE) with 28.96ms inference time, suggesting excessive feature injection introduces redundancy without improving accuracy-latency trade-offs. The superior performance of medium-scale architectures paired with large vision transformers suggests an optimal balance: M-ViT-B-Dual achieves competitive accuracy (56.18% Tunnel, 67.89% KITTI, 52.39% PPE) with the most favorable inference time (21.18ms) among DINO-YOLO variants, enabling real-time deployment at 47 FPS for edge computing scenarios.

The baseline YOLOv12 performance exhibits notable inconsistencies highlighting DINO-YOLO's value. YOLOv12-M achieves highest baseline performance on Tunnel (53.42%) and PPE (49.04%), while YOLOv12-S leads on KITTI (41.54%), and YOLOv12-L shows poorest performance on KITTI (36.21%) despite being the largest architecture. This counterintuitive scaling behavior indicates conventional supervised learning suffers from overfitting in data-constrained regimes, where increased capacity without adequate training data degrades generalization. In contrast, DINO-YOLO variants consistently benefit from larger architectures, with L-ViT-B-Dual (72.06%) substantially outperforming S-ViT-B-Triple (68.16%) on KITTI, validating that self-supervised pre-training fundamentally changes data efficiency characteristics of large models.

The relative improvement patterns provide empirical validation of transfer learning effectiveness as a function of dataset size. KITTI's 73.5% relative improvement represents the inflection point where self-supervised features provide transformative benefits, PPE's 13.7% improvement shows significant but not transformative benefits, and Tunnel's 5.2% improvement suggests diminishing returns in extreme scarcity where even powerful pre-trained features struggle with limited supervision and domain-specific complexity. This characterization informs practical deployment decisions: for moderate data regimes (1-10K images), DINO-YOLO provides substantial returns; for extreme scarcity (<1K images), complementary strategies such as synthetic data generation or active learning become necessary.

The computational overhead of DINO-YOLO variants (21-33ms inference time versus 8-16ms for baselines) represents acceptable trade-offs for civil engineering applications. The 2-4× latency increase maintains real-time processing capability (30-47 FPS) sufficient for construction site monitoring, autonomous equipment navigation, and quality control inspection workflows. For applications requiring maximum throughput, M-ViT-B-Dual achieves optimal efficiency with 21.18ms inference time while delivering strong performance across all datasets, enabling deployment on mid-range GPUs (NVIDIA RTX 5090) or edge devices (Jetson AGX Orin) where computational budgets constrain model selection. For applications prioritizing maximum accuracy where computational resources permit, L-ViT-B-Dual and M-ViT-L-Dual justify their higher inference costs (33.25ms and 28.09ms respectively) through substantial performance improvements, particularly on KITTI where 72.06% mAP@0.5 enables semi-autonomous construction vehicle assistance systems with acceptable safety margins.

5. Ablation study

The systematic ablation study in Table 3 reveals critical insights into DINO-YOLO architectural design choices, demonstrating that optimal integration strategies depend strongly on YOLO backbone scale, DINOv3 model capacity, and feature injection locations. The experimental results expose complex interactions between these architectural dimensions that fundamentally determine detection accuracy in data-constrained construction safety monitoring applications.

The ablation results demonstrate that optimal DINO-YOLO configurations vary systematically across YOLO scales, contradicting the hypothesis that a single universal integration strategy exists for all model sizes. Large-scale architectures achieve best performance with L-vitb16-dualp0p3 (53.08% mAP@0.5, +16.9% improvement), combining the moderately-sized ViT-B/16 transformer with DualP0P3 integration at input preprocessing (P0) and mid-backbone (P3). This configuration provides complementary semantic enhancement: P0 transforms raw pixels into semantically-enriched inputs benefiting all subsequent convolutional layers, while P3 directly enhances the 80×80 feature maps feeding the feature pyramid network for multi-scale detection. The Large backbone's substantial capacity (approximately 90M parameters in detection-specific layers) provides sufficient representational power to effectively integrate semantic features from two DINOv3 injection points without suffering architectural interference or gradient flow complications.

Medium-scale architectures achieve optimal performance with M-vitl16-dualp0p3 (55.77%, +13.7%), notably employing the larger ViT-L/16 transformer rather than ViT-B/16. This counterintuitive finding indicates that semantic feature quality dominates architectural capacity matching the 307M-parameter ViT-L/16 captures more nuanced visual patterns from self-supervised pre-training than the 86M-parameter ViT-B/16, and these superior semantic representations compensate for the Medium backbone's reduced detection capacity. The M-vitl16-dualp0p3 configuration's superiority over both baseline and all other Medium-scale variants validates that strategic integration of high-quality pre-trained features enables smaller detection backbones to achieve performance exceeding larger backbones with inferior feature representations.

Small-scale architectures exhibit fundamentally different optimal configuration patterns, with S-vitb16-triple achieving highest performance (53.63%, +10.5%) using Triple Integration (P0-P3-P4)—a strategy that catastrophically degrades Large-scale performance (L-vitb16-triple: 28.78%, -36.6%). The Small backbone's success with triple integration can be explained through limited capacity considerations: with fewer parameters in detection-specific layers (approximately 25M), the Small architecture benefits from maximal semantic guidance across multiple hierarchical levels, where DINOv3 features effectively compensate for the backbone's limited capacity to learn complex representations from restricted training data. However, the Small-scale architecture cannot effectively leverage the larger ViT-L/16 model, with S-vitl16-dual achieving only 38.56%—a 20.6% degradation indicating severe architectural mismatch where the Small backbone's limited capacity cannot integrate the rich 307M-parameter ViT-L/16 representations.

The DualP0P3 integration strategy demonstrates consistent effectiveness across Medium and Large scales when paired with appropriately-sized DINO variants, with L-vitb16-dualp0p3 (+16.9%) and M-vitl16-dualp0p3 (+13.7%) achieving highest performance within their respective scale categories. However, this same strategy produces performance collapse on Small architectures, with S-vitb16-dualp0p3 achieving only 41.92% (-13.6% degradation), suggesting Small backbones lack sufficient capacity to simultaneously process semantic features from two injection points. This finding has critical implications for edge deployment: practitioners using Small architectures on resource-constrained devices (NVIDIA Jetson, mobile platforms) should avoid DualP0P3 despite its effectiveness on larger models, accepting this capacity constraint as fundamental to maintaining detection pipeline stability.

The catastrophic failure of L-vitb16-triple (28.78%, representing the worst performance across all ablation experiments) demonstrates that excessive feature injection can fundamentally destabilize detection pipelines even in high-capacity architectures. Triple Integration injects DINOv3 features at three hierarchical levels (P0 at 640×640 resolution, P3 at 80×80, P4 at 40×40), creating redundant semantic signals that interfere with the backbone's learned feature abstractions. The performance collapse suggests gradient flow complications where backpropagation signals from three frozen DINOv3 branches create conflicting optimization directions, representational redundancy where semantically similar features at multiple scales provide diminishing marginal information, or architectural interference disrupting YOLO's carefully-tuned multi-scale feature pyramid hierarchy. The fact that this same Triple Integration strategy achieves optimal Small-scale performance indicates failure arises from architectural mismatch rather than fundamental strategy deficiencies—larger backbones possess sufficient internal feature learning capacity that makes extensive external feature injection counterproductive.

Single-integration configurations provide insights into relative importance of different injection locations, consistently producing modest improvements (+1-3%) or slight degradations substantially underperforming optimal multi-point configurations. The superiority of P0 single integration over P3 injection (M-vitl16-singleP0: 50.51% versus M-vitl16-singleP3: 46.45%) can be explained through information propagation mechanics: features injected at input level influence all subsequent convolutional layers through entire backbone depth, while P3-level injection only affects downstream layers. This suggests practitioners limited to single-integration configurations should prioritize P0 injection to maximize performance, though the relatively modest improvements compared to optimal multi-point configurations (+10-17%) indicate practitioners should invest in dual or triple integration whenever deployment constraints permit.

The relationship between DINO model capacity and detection performance depends critically on YOLO backbone scale, revealing complex architectural matching requirements. For Medium-scale architectures, larger DINO models consistently improve performance: M-vitl16-dualp0p3 (55.77%) substantially outperforms M-vitb16-dualp0p3 (52.39%), representing a 6.4% relative improvement attributable solely to upgrading from ViT-B/16 to ViT-L/16. This positive scaling validates that richer semantic representations from larger vision transformers transfer more effectively to detection tasks. However, this relationship inverts dramatically for Small-scale architectures, where S-vitl16-dual achieves only 38.56%—a catastrophic failure demonstrating that architectural capacity matching represents a critical constraint. Large-scale architectures exhibit intermediate behavior where L-vitl16-dualp0p3 (50.84%) performs slightly worse than L-vitb16-dualp0p3 (53.08%), indicating optimal performance with moderately-sized transformers rather than maximum-capacity models and suggesting architectural optimization requires balancing semantic representation quality against integration complexity.

Based on these ablation findings, practitioners should select configurations according to deployment requirements and constraints. For applications prioritizing maximum detection accuracy where computational resources permit Large or Medium-scale backbones, optimal configurations are M-vitl16-dualp0p3 (55.77%) or L-vitb16-dualp0p3 (53.08%), with the Medium configuration's superior absolute performance despite smaller backbone size validating that high-quality DINO features combined with strategic integration enable efficient architectures to outperform larger models. For resource-constrained edge deployment scenarios requiring Small-scale architectures—embedded systems on construction vehicles, battery-powered mobile robots, or drone-based surveys—practitioners should deploy S-vitb16-triple (53.63%) while strictly avoiding ViT-L/16 variants that cause catastrophic performance degradation, though if computational overhead exceeds device constraints, S-vitb16-dual (50.90%) provides reasonable compromise with reduced demands. For practitioners facing uncertainty about deployment scale requirements or seeking robust configurations that generalize across varying hardware platforms, DualP0P3 integration with ViT-B/16 represents the safest default choice, achieving strong Large-scale performance (53.08%), reasonable Medium-scale performance (52.39%), and avoiding catastrophic failure modes, providing a principled architecture selection strategy grounded in complementary information propagation mechanisms rather than exhaustive trial-and-error hyperparameter search.

6. Visualization

The visualization presents feature map (Fig. 5) activations from convolutional layers positioned after the DINO integration module in a YOLO-L (Large variant) architecture applied to autonomous driving scenes from the KITTI dataset, comparing implementations with and without DINO feature incorporation using pseudocolor representation where activation intensities range from purple (low) through cyan and green to yellow (high). The DINO-enhanced configuration exhibits fundamentally transformed convolutional feature representations: (1) significantly amplified edge responses with intense cyan and green activations along building facades, vehicle contours, and road lane markings, demonstrating effective leveraging of transformer-derived geometric priors for structural feature extraction; (2) spatially coherent contextual activation patterns forming connected distributions across semantically related regions such as building complexes and vehicle clusters, reflecting scene-level relationship understanding critical for autonomous driving; (3) robust foreground-background segregation with pronounced contrast between drivable road surfaces (purple regions) and obstacle elements including vehicles and pedestrians (cyan/green/yellow regions); (4) expanded dynamic range with high-intensity yellow activations on safety-critical objects such as nearby vehicles and pedestrians; (5) channel-specific semantic specialization with selective responses to distinct KITTI object categories (cars, pedestrians, cyclists, infrastructure); and (6) depth-aware activation patterns where intensity correlates with object proximity. Conversely, the baseline YOLO-L without DINO exhibits spatially constrained, fragmented activation patterns; substantially compressed dynamic range dominated by purple coloration indicating weak feature responses; absence of structured hierarchical progression; diffuse non-discriminative activations failing to establish clear semantic boundaries; and minimal sensitivity to geometric structures essential for road scene parsing.

7. Discussion

This research introduces DINO-YOLO, a hybrid architecture integrating DINOv3 self-supervised vision transformers with YOLOv12 for data-efficient object detection in civil engineering applications. Evaluation across datasets spanning 648 to 118,000 images demonstrates that self-supervised pre-training provides substantial benefits in moderate data regimes while maintaining real-time inference capability (30-47 FPS).

Key Findings:

DINO-YOLO achieves significant performance improvements over baseline architectures, ranging from 12.4% in extreme scarcity (648 images) to 88.6% in moderate regimes (5,233 images). The KITTI dataset results—72.06% mAP@0.5 versus 38.21% baseline—establish state-of-the-art performance for civil engineering datasets with 1,000-10,000 images. Systematic ablation reveals optimal configurations vary by scale: Medium-scale architectures perform best with DualP0P3 using ViT-L/16 (55.77% mAP@0.5), Large-scale with DualP0P3 using ViT-B/16 (53.08%), and Small-scale with Triple Integration using ViT-B/16 (53.63%). On COCO, DINO-YOLO achieves Pareto-optimal efficiency (53.5% mAP@0.5:0.95 with 25-30M parameters versus 40-80M for comparable architectures).

Practical Implications:

DINO-YOLO enables deployment with 5,000-10,000 annotated images—an order of magnitude reduction compared to conventional requirements. The 2-4× inference overhead (21-33ms versus 8-16ms) maintains real-time processing while reducing GPU memory from 24-32GB to 8-12GB, enabling mid-range hardware deployment and reducing infrastructure costs from $50,000+ to $15,000-20,000 for multi-camera installations. Configuration guidance: M-ViT-B-Dual (21.18ms) optimizes edge deployment, M-ViT-L-Dual (28.09ms) balances accuracy for multi-camera systems, and L-ViT-B-Dual (33.25ms) maximizes performance where resources permit.

Limitations and Future Directions:

Performance in extreme scarcity regimes (<1,000 images) remains limited, with tunnel segment crack detection achieving 12.4% improvement and 56.18% absolute performance. The 56.18% ceiling reflects task complexity rather than architectural deficiency. Future work should explore: (1) active learning frameworks for 40-60% annotation reduction, (2) physics-informed architectures incorporating domain-specific knowledge, (3) synthetic data generation, (4) multi-modal sensor fusion, and (5) hierarchical domain adaptation through construction-specific self-supervised learning.

DINO-YOLO establishes state-of-the-art performance for specialized civil engineering datasets (<10,000 images) while preserving computational efficiency for field deployment, providing practical solutions for construction safety monitoring, infrastructure inspection, and automated quality control in data-constrained environments.

8. Conclusion

This research introduces DINO-YOLO, a hybrid architecture that integrates DINOv3 self-supervised vision transformers with YOLOv12 for data-efficient object detection in civil engineering applications. Systematic evaluation across datasets spanning 648 to 118,000 images demonstrates that self-supervised pre-training provides substantial performance benefits in moderate data regimes while maintaining real-time inference capability (30-47 FPS).

DINO-YOLO achieves significant improvements over baseline architectures across all evaluated datasets. In extreme scarcity regimes (648 images), performance gains of 12.4% were observed. The most substantial improvements occurred in moderate data regimes, with the KITTI dataset (5,233 images) achieving 72.06% mAP@0.5 compared to the baseline 38.21%, representing an 88.6% relative improvement. This establishes state-of-the-art performance for civil engineering datasets containing 1,000-10,000 images.

Systematic ablation studies reveal that optimal configurations vary according to model scale. Medium-scale architectures achieve best performance with DualP0P3 integration using ViT-L/16 (55.77% mAP@0.5). Large-scale architectures perform optimally with DualP0P3 using ViT-B/16 (53.08%), while Small-scale architectures require Triple Integration using ViT-B/16 (53.63%). On the COCO dataset, DINO-YOLO demonstrates Pareto-optimal efficiency, achieving 53.5% mAP@0.5:0.95 with 25-30M parameters compared to 40-80M for comparable architectures.

DINO-YOLO enables practical deployment with 5,000-10,000 annotated images, representing an order of magnitude reduction compared to conventional supervised learning requirements. The architecture incurs a 2-4× inference overhead (21-33ms versus 8-16ms baseline) while reducing GPU memory requirements from 24-32GB to 8-12GB. This enables deployment on mid-range hardware (NVIDIA RTX 5090, Tesla T4) rather than requiring high-end accelerators for multi-camera installations.

Three deployment configurations are recommended: M-ViT-B-Dual (21.18ms, 47 FPS) optimizes edge deployment and multi-camera systems; M-ViT-L-Dual (28.09ms, 36 FPS) provides enhanced accuracy for complex scenarios; and L-ViT-B-Dual (33.25ms, 30 FPS) maximizes performance where computational resources are available.

Performance in extreme scarcity regimes (<1,000 images) remains limited. Tunnel segment crack detection achieved only 12.4% improvement with 56.18% absolute performance, reflecting task complexity rather than architectural deficiency.

In conclusion, DINO-YOLO establishes state-of-the-art performance for specialized civil engineering datasets (<10,000 images) while preserving computational efficiency for field deployment. The framework provides practical solutions for construction safety monitoring, infrastructure inspection, and automated quality control in data-constrained operational environments.