RWDS-CZ Domain Shift Project: Key Code

This folder is a lightweight code snapshot for the CS2756 final report. It is not a full training repository. Large datasets, generated images, feature caches, model checkpoints, SAM3 semantic maps, and object cutout banks are intentionally excluded.

Contents

• domain_analysis/scripts/: feature extraction, MMD/Proxy A-distance, PCA/ANOVA, scene-vs-object MMD, and follow-up checks used in the report.
• semantic_copy_paste/: the MMDetection transform for online long-tail copy-paste, plus scripts for CLIP indexing, SAM3 semantic maps, and object-background co-occurrence.
• synthetic_alignment/: FLUX Kontext climate-transfer scripts and the 1024-to-512 COCO repackaging utilities used for synthetic alignment.
• mmdet_configs/: the small set of Grounding-DINO configs corresponding to the baseline, plain copy-paste, semantic copy-paste, and B2C/C2B synthetic alignment experiments.

Data Paths

The domain-analysis scripts read RWDS-CZ from RWDS_CZ_ROOT.

export RWDS_CZ_ROOT=/path/to/RWDS_Dataset/RWDS_CZ
cd domain_analysis/scripts
PYTHON=python bash run_all.sh

The copy-paste configs expect the usual MMDetection layout with RWDS_Dataset/RWDS_CZ, rwds_copypaste/assets, and object cutout directories available beside the MMDetection checkout. The key algorithmic file is semantic_copy_paste/transforms/online_copy_paste.py.

The synthetic-alignment scripts use RWDS_ROOT and RWDS_OUTPUT_ROOT, or the matching command-line flags:

python synthetic_alignment/generation/run_kontext_group_transfer_fast.py \
  --rwds-root /path/to/RWDS_Dataset/RWDS_CZ \
  --output-root /path/to/RWDS_Dataset_Kontext/RWDS_CZ

External Dependencies

These files rely on the original project environments: PyTorch, transformers, open_clip, timm, scikit-learn, scipy, matplotlib, pandas, MMDetection, FLUX/Kontext dependencies, and SAM3 for semantic maps.

The submitted code focuses on the methods and analysis used in the report; it omits heavyweight assets so the GitHub repository remains small.