PariSolunke/africa-seg-bench

# Africa Segmentation Benchmarks Satellite segmentation benchmark pipeline. ## Layout ``` data/ → images/, gray/, annotations/ per split train|val|test|whole_data/ images/ RGB satellite tiles gray/ ground-truth class-id masks annotations/ RGB color masks for visualization configs/ classes.yaml → class ids, colors, SAM3 prompts dataset.yaml → split paths inference/*.yaml → per-model configs results/ → predictions and metrics per run {run_name}/{split}/ masks_gray/ masks_rgb/ metrics/ scripts/ run_inference.py → run any model on a split run_metrics.py → compare predictions against ground truth slurm/ → HPC batch scripts src/africa_benchmark/ inference/adapters/ → one file per model (sam3_concept.py, gemini.py, ...) ``` ## How to ### 1. Install dependencies into a Singularity overlay (HPC) ```bash # SAM3 needs an overlay with torch 2.10 + CUDA 12.8 - sbatch --export=ALL,INSTALL_SAM3=1,INSTALL_SAM2=0 scripts/slurm/install_to_overlay.sbatch ``` ### 2. Run benchmarks ```bash # SAM3 sbatch --export=ALL,MODEL_CONFIG=configs/inference/sam3_concept.yaml,RUN_NAME=sam3_concept,SPLITS="val test",HF_TOKEN="YOUR_TOKEN_HERE" scripts/slurm/run_split_benchmark.sbatch # Gemini sbatch --export=ALL,MODEL_CONFIG=configs/inference/gemini.yaml,RUN_NAME=gemini,SPLITS="val test",GEMINI_API_KEY="YOUR_API_KEY_HERE" scripts/slurm/run_split_benchmark.sbatch ``` Results land in `results/{run_name}/{split}/masks_gray/`. Metrics are under `results/{run_name}/{split}/metrics/`.