Benchmark method adapters

This directory intentionally contains only RAD-owned adapters, dataset conversion code, configuration, and prompts. It does not vendor full third-party repositories or CUDA submodules. Install each upstream at the exact revision in manifest.toml, then use the corresponding adapter here. This keeps provenance, updates, and license terms visible.

Paper family	Baselines	RAD-specific code	What changed for RAD
2D feature	CFlow, EfficientAD, FastFlow, PaDiM, PatchCore, Reverse Distillation, STFPM, UFlow, WinCLIP	src/rad_benchmark/anomalib.py	Maps each physical instance's MVTec-style folders into Anomalib Folder, keeps normal test images separate, pairs every defect directory with its mask directory, and skips fitting WinCLIP.
Zero-shot CLIP	AdaCLIP	adaclip/rad_dataset.py	Loads RAD path metadata without hard-coded roots and preserves semantic-category plus physical-instance labels.
Zero-shot CLIP	VCP-CLIP	vcpclip/convert_rad.py	Converts RAD into VCP-CLIP's unified good/anomaly layout with collision-free names and binary masks.
3D reconstruction	SplatPose, PIAD	reconstruction/prepare_splatpose.py, reconstruction/run.py	Rewrites camera-frame paths in a separate prepared tree and constructs an explicit SplatPose command. PIAD stops at portable config generation because its method source is not public.
VLM	Qwen2.5-VL, GPT-4o	vlm/prompt.json, src/rad_benchmark/vlm.py	Enforces a provider-neutral JSON response, converts predicted boxes to binary masks, and never stores credentials.

Fetch upstreams

The helper clones to an ignored third_party/ directory and checks out pinned revisions:

bash scripts/fetch_upstream.sh anomalib
bash scripts/fetch_upstream.sh adaclip
bash scripts/fetch_upstream.sh vcpclip
bash scripts/fetch_upstream.sh splatpose

Review the upstream license before downloading weights or running a method. In particular, SplatPose's Gaussian Splatting components are not covered by RAD's MIT license and may restrict use or redistribution.

Smoke checks

These commands do not load a model or dataset:

python -m rad_benchmark.cli --help
python methods/vcpclip/convert_rad.py --help
python methods/reconstruction/prepare_splatpose.py --help
python methods/reconstruction/run.py --help
bash scripts/fetch_upstream.sh --help
bash scripts/run_clip.sh --help

Full experiments require method-specific CUDA environments and checkpoints. PIAD source is not public as of this release, so its adapter stops after generating a portable config. The paper uses COLMAP estimates from unmasked images for SplatPose and PIAD; robot ground-truth poses are metadata, not the reported 3D-baseline input.