Projects

Production road-segmentation engine built on Mask2Former with a Swin-L backbone — deployed and running live on NVIDIA V100, and serving as the supervised reference for the DINOv3 system.

• 76.86 IoU on pothole-class segmentation (held-out production validation set) — the product metric that governs how reliably a road defect is found and measured in the field.
• 60.89 mIoU zero-shot on 152 never-seen Australian urban street images across 15 classes, including night scenes — evidence the model learned roads, not a dataset.
• 1.05 s single forward pass on the production V100; 1.8 s per image for the full 6-vote test-time-augmentation pipeline behind the headline result.
• Dual-head design with road-gated false-positive filtering; Swin-L hierarchical transformer (ImageNet-22K pretrained) with Mask2Former's masked and multi-scale deformable attention.
• Same training infrastructure and data pipeline as the DINOv3 variant, enabling fair supervised-vs-self-supervised architecture comparisons.

State-of-the-art road segmentation system powered by DINOv3 ViT-L backbone and Mask2Former head, trained on large-scale geospatial datasets for production deployment.

• DINOv3 backbone: self-supervised pretrained ViT-L with 300M+ parameters, fine-tuned for road surface understanding across diverse conditions.
• Custom multi-phase training pipeline: progressive resolution scaling (512→1024→1536), mixed precision (FP16), gradient accumulation for large effective batch sizes.
• Custom loss design: a boundary-aware loss for fine-grained edge preservation and a cross-scale consistency loss, combined with standard Dice + CrossEntropy.
• Custom attention and gating modules for road-specific feature refinement.
• Automated checkpoint management and seamless training resumption for long-running experiments (80K+ iterations).

Real-time object detection system for road assets and infrastructure, built on the latest YOLO architecture.

• YOLO26 architecture: latest anchor-free detection paradigm with advanced feature pyramid and decoupled head design.
• Optimised for real-time inference on edge devices and cloud GPU instances.
• Custom dataset curation and augmentation pipeline for road-specific object categories.
• Integrated with the shared ML infrastructure: automated evaluation, metric tracking, and model versioning.