IRAOD is an oriented object detection workspace for DIOR/RSAR source-free adaptation and corruption robustness experiments.
bash scripts/setup_env_iraod.sh
conda activate iraod
python tools/verify_dataset_layout.py --dior dataset/DIOR --rsar dataset/RSAR
DRY_RUN=1 SOURCE_CKPT=auto WORK_ROOT=work_dirs/smoke_refactor \
CUDA_VISIBLE_DEVICES=0,1,2 NGPUS=3 bash scripts/run/rsar_sfodrs_full.shUse DRY_RUN=1 first on any launcher to verify paths, env, and generated
commands without starting training or evaluation.
# Full SFOD-RS protocol. SOURCE_CKPT=auto trains the clean source detector first.
SOURCE_CKPT=auto WORK_ROOT=work_dirs/<run> CUDA_VISIBLE_DEVICES=0,1,2 NGPUS=3 \
bash scripts/run/rsar_sfodrs_full.sh
# One corruption domain.
CORR=chaff SOURCE_CKPT=work_dirs/<source>/latest.pth WORK_ROOT=work_dirs/<run> \
bash scripts/run/rsar_sfodrs_domain.sh
# TENT adapt/eval/fusion.
TENT_MODE=adapt SOURCE_CKPT=work_dirs/<source>/latest.pth WORK_ROOT=work_dirs/<tent> \
bash scripts/run/rsar_tent.sh
# Collect SFOD-RS result tables.
WORK_ROOT=work_dirs/<run> bash scripts/run/collect_results.shCompatibility wrappers remain for older commands such as:
bash scripts/run_rsar_sfodrs_full_3gpu.sh auto work_dirs/<run>
bash scripts/run_rsar_sfodrs_7corr.sh work_dirs/<source>/latest.pth
bash scripts/exp_rsar_sfodrs_adapt.sh chaff work_dirs/<source>/latest.pthDefault local layout:
dataset/DIOR/
dataset/RSAR/
train/{images,annfiles}
val/{images,annfiles}
test/{images,annfiles}
corruptions/<corruption>/{val,test}/images
Override RSAR location with either RSAR_DATA_ROOT=/path/to/RSAR or
DATA_ROOT=/path/to/RSAR.
Prepare the RSAR corruption layout when needed:
python tools/prepare_rsar_corruption.py --data-root dataset/RSAR --workers 8docs/commands.md: current smoke/full/eval/collect command cookbook.docs/project_map.md: directory responsibilities and active vs archived entrypoints.docs/experiment.md: full experiment ledger and recorded metrics.configs/README.md: current config aliases and historical config locations.MODEL_ZOO.md: checkpoint and external weight notes.