Visual Pollution Segmentation

Segmentation of urban visual pollution (9 classes + background: billboard, banner, poster, neon box, shop signage, graffiti, roadside ad, utility/overhead cable, digital screen/LED) and distillation of SAM 3 into small on-device models (SHSSNet).

• Code, reports, paper → this repo
• Model weights → Hugging Face: <hf-user>/visual-pollution-checkpoints (see huggingface_models/MANIFEST.csv for run → file mapping)
• Datasets / teacher caches → not stored here (datasets are private; teacher caches are regenerable with the SAM 3 scripts)

Layout

Folder	What it is
experiments/visual_pollution_seg_gdino_mobilesam	Early GroundingDINO + MobileSAM pseudo-labelling, SegFormer-B0
experiments/visual_pollution_segmentation	SAM 3 pseudo-labels → SegFormer-B0 pipeline
experiments/billboard_sam3_pseudo_masks	SAM 3 pseudo masks for the billboard dataset
experiments/billboard_seg_baseline	PIDNet-S / SegFormer-B1 baselines
experiments/model_comparison_5models, segmentation_model_benchmark	Early model comparisons
experiments/metric_depth_benchmark	Metric depth model comparison
experiments/shssnet_distillation	SHSSNet distillation from SAM 3 (variants C / C-wide / P, ablations, grid)
experiments/shssnet_kd_sam3	SHSSNet KD with SAM 3 teacher (Kaggle runs)
experiments/shssnet_ds5_benchmark	11-checkpoint benchmark on seg_ds5_pakEvert (outputs/reports/REPORT.md)
experiments/dari_*	Result logs/reports pulled back from Kaggle / Windows training runs
experiments/shssnet_report	LaTeX report (SHSSNet_Report.tex/.pdf, build.sh)
paper/	Architecture design doc, conference paper (main.pdf)

Each experiment folder keeps its original name so relative cross-references between sibling folders (e.g. the ds5 benchmark reading shssnet_distillation) still resolve. Scripts that expect checkpoints at outputs/checkpoints/<run>/best.pt need the weights downloaded from Hugging Face first (see huggingface_models/README.md).