1. Novel Hybrid Deep Learning Architecture

Engineered and implemented an end-to-end instance segmentation network integrating Mask R-CNN with Deep Metric Learning, replacing the standard parametric linear classification layer with a Squared Euclidean Prototypical Predictor head.

2. Tackling Extreme Class Imbalance

Formulated a geometric centroid-learning framework to resolve the severe long-tail distribution and gradient dominance inherent in vehicle exterior inspection datasets (CarDD benchmark), eliminating minority class collapse (e.g. flat tires, shattered glass).

3. Hyperspherical Latent Geometry

Enforced strict $L_2$ normalisation on both 1024-D RoI feature embeddings and learnable class prototypes, projecting representations onto a unit hypersphere ($\|p\| = 1$). Mathematically derived that pairwise squared Euclidean distance:

Proved mathematically and empirically that the model converged to an orthogonal feature space that effectively eliminates inter-class visual confusion.

4. Optimization & Mixed-Precision Protocol

Leveraged a ResNet-50-FPN backbone pre-trained on MS-COCO, configured with temperature-scaled softmax ($\alpha = 20$) for rapid monotonic loss convergence, SGD with momentum, and learning rate step-decay. Utilised Automatic Mixed Precision (torch.cuda.amp) to maximize memory efficiency on an NVIDIA RTX 3060 (12GB).

5. Data Augmentation & Spatial Fidelity

Engineered synchronized spatial transformations (torchvision.transforms) dynamically mirroring binary segmentation masks and bounding box coordinates during horizontal flips to maintain pixel-level spatial fidelity under varying sensor orientations.

6. Future Vision-Language-Action (VLA) Roadmap

Proposed extending metric damage embeddings into the latent space of Multimodal Large Language Models (MLLMs) and VLA frameworks for automated, explainable vehicle safety diagnostics in autonomous mobility ecosystems.