Flood Depth Estimation on Flood Depth dataset (test)
5.65MAE (cm)FloodLlama
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| FloodLlamaTraining data=340,416 samples (single vehicle), Dataset split=50/50 scene-level train/test, Epochs=3, Learning rate=2×10⁻⁴, LR schedule=Cosine annealing (10 warm-up steps), Per-device batch size=2, Gradient accumulation=4, Effective batch size=8, Steps per epoch=3,750, Optimizer=AdamW (Paged), Hardware=NVIDIA RTX 4090 (24 GB VRAM), Training duration=10.3 hours, Output checkpoint=llama32_flood_depth_ckpt_full2ep2026.03 | 5.65 | 7.38 | 23.23 | 75.42 |