Problem

Fine-grained object detection in challenging domains like vehicle damage assessment is difficult, even for human experts. Existing methods, such as DiffusionDet, struggle with subtle, low-contrast damage like scratches and cracks due to their reliance on local feature conditioning. These local features can be confounded by visual clutter, reflections, and varied lighting, making accurate localization and classification challenging. The CarDD dataset highlights these limitations, showing that current detectors often miss faint damages or produce imprecise bounding boxes, particularly for small objects, because they lack a comprehensive understanding of the entire scene.

Method

C-DiffDet+ addresses the limitations of local feature conditioning in diffusion-based detectors by integrating global scene context. The framework builds upon the DiffusionDet architecture and introduces three key components: an Adaptive Channel Enhancement (ACE) block, a Global Context Encoder (GCE), and a Context-Aware Fusion (CAF) module. The ACE blocks refine backbone and Feature Pyramid Network (FPN) features, enhancing their discriminative power. The GCE extracts a compact embedding of the entire scene, capturing holistic environmental information through a residual CNN encoder. The CAF module then uses cross-attention mechanisms to seamlessly integrate this global scene context with the local Region of Interest (RoI) features from object proposals. Additionally, an enhanced Multi-Modal Fusion (MMF) module incorporates global context embeddings along with temporal and positional embeddings, creating a unified representation for iterative refinement. This allows each object proposal to attend to comprehensive scene-level understanding during the denoising diffusion process, leading to more robust and accurate predictions.

Results

C-DiffDet+ significantly outperforms state-of-the-art models on the CarDD benchmark, achieving a remarkable APbb of 64.8%, a 1.4% improvement over DiffusionDet. The model shows substantial gains in mean Average Precision, especially for small object detection (APbbS of 45.5%, a 6.8% increase) and at stricter IoU thresholds (APbb75 of 67.9%, a 1.7% increase). In per-category analysis, C-DiffDet+ establishes new state-of-the-art performance for crack detection (42.2% AP, +7.1%), lamp broken (80.2% AP, +3.4%), and glass shatter (94.2% AP). An ablation study confirms the synergistic effects of ACE, GCE, and CAF, demonstrating that their combined use leads to optimal performance. The model also shows improved generalization on the VehiDE dataset, achieving an overall AP of 33.9%, a 0.6% improvement over DiffusionDet, particularly in localization precision at higher IoU thresholds (AP75 of 34.1%). Qualitative visual comparisons further illustrate that C-DiffDet+ produces more precise and contour-adherent bounding boxes, effectively detecting faint and low-contrast damage that other methods miss, with increased prediction confidence.

Implications

The introduction of C-DiffDet+ and its context-aware fusion mechanisms represents a significant advancement in fine-grained object detection, particularly for automotive damage assessment. By effectively integrating global scene context with local features in a diffusion-based framework, the model overcomes critical limitations of previous methods, leading to more accurate and robust detection of subtle and challenging damage types like scratches and cracks. This improved capability has direct implications for real-world applications such as automated vehicle inspection and insurance claims processing, where high-fidelity damage detection is crucial. The methodology also opens new avenues for research into more robust loss functions that are less sensitive to annotation noise, and alternative damage representations like segmentation masks or polygons for better handling of amorphous damage shapes. Future work will focus on improving efficiency for real-time deployment and generalizing this context-aware paradigm to other vision tasks requiring global scene understanding, potentially through self-supervised pre-training.