HUST-CVLab-PS/lino-lvc-checkpoints

LiNo-UniPS + LVC checkpoints

Checkpoints for bachelor thesis: Lighting Variation Confidence (LVC) module on LiNo-UniPS — Đoàn Anh Vũ, HUST SOICT.

Variants

Folder	Variant	Description
bl_real_*	baseline	Original LiNo-UniPS (no LVC)
fl_real_*	lvc_full	LiNo-UniPS + LVC full (loss weighting + feature scaling, proposed method)
ll_real_*	lvc_loss	LVC loss weighting only (ablation)
lf_real_*	lvc_feat	LVC feature scaling only (ablation)

Filename convention: {variant_short}_{mode}_{DDMMYYYY}_{HHMMSS}_e{epoch}_v{val_loss}.ckpt

• variant_short: bl / fl / ll / lf
• mode: smoke (test) / real (production)

Training setup

• Dataset: HDLong + (PolarPS future), hosted at HUST-CVLab-PS/UniPS
• Hardware: NVIDIA RTX 6000 Ada (Vast.ai)
• Optimizer: AdamW(lr=1e-4, wd=0.05), StepLR(step=10, gamma=0.8)
• Precision: bf16-mixed
• K input images: 6 (HDLong)
• Hyperparams khác xem wandb run config

Acknowledgments

Base architecture & weights: LiNo-UniPS (MIT License) by H. Li, H. Chen et al.

• Paper: arXiv 2506.18882
• Original model card: houyuanchen/lino

License

MIT — same as upstream LiNo-UniPS.