LoGo-Fuse

LoGo-Fuse is a training-free 3D point cloud OOD detection codebase built on frozen ULIP-2 features.

This public repository keeps the cleaned mainline only:

• local graph propagation
• global prototype refinement
• negative prototype bank
• scalar local/global fusion

Historical experimental branches such as gsp_ot, old OT/OODD scoring, geometry-side fusion branches, and legacy few-shot prototype clustering variants are not part of the intended public workflow.

Repository Contents

Included:

• main_logofuse.py
• ood_methods/, models/, utils/
• data/ metadata and dataset configs
• tools/ helper scripts

Not included in GitHub:

• checkpoints (*.pt, *.pth)
• dataset binaries (*.dat)
• runtime caches and logs (outputs/, logs_*)

Environment

Python 3.10+ is recommended.

CUDA is required for ULIP-2 evaluation in this repository.

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

Install PyTorch separately for your CUDA version. Example for CUDA 11.8:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Optional acceleration for PointBERT KNN:

pip install --no-build-isolation git+https://github.com/unlimblue/KNN_CUDA.git

If knn_cuda is not installed, LoGo-Fuse falls back to the built-in torch KNN implementation.

Required Assets

Checkpoint

Download the ULIP-2 checkpoint:

• ULIP-2 checkpoint

Place it at the repository root as:

ULIP-2-PointBERT-8k-xyz-pc-slip_vit_b-objaverse-pretrained.pt

or override with:

export TEST_CKPT_ADDR=/abs/path/to/ULIP-2-PointBERT-8k-xyz-pc-slip_vit_b-objaverse-pretrained.pt

Dataset .dat bundle

Download the dat bundle:

• LoGo-Fuse DAT bundle

Expected paths:

data/scanobjectnn15_normal_resampled/scanobjectnn15_train_2048pts_fps.dat
data/scanobjectnn15_normal_resampled/scanobjectnn15_test_2048pts_fps.dat

data/shapenetcore54_normal_resampled/shapenetcore54_train_4096pts_fps.dat
data/shapenetcore54_normal_resampled/shapenetcore54_test_4096pts_fps.dat

data/modelnet40_normal_resampled/modelnet40_train_8192pts_fps.dat
data/modelnet40_normal_resampled/modelnet40_test_8192pts_fps.dat

The repository already includes the required lightweight metadata such as:

• data/templates.json
• data/labels.json
• data/SR, data/SN, data/MN
• *_train.txt, *_test.txt, *_shape_names.txt

Benchmark Tracks

This repository currently documents three evaluation tracks.

1. Synthetic: ShapeNetCore54

Zero-shot example (SN1):

python main_logofuse.py \
  --model ULIP_PointBERT --method logofuse --evaluate_3d \
  --dataset_name ShapeNetCore54 --dataset_split SN1 \
  --npoints 4096 --validate_dataset_prompt shapenet_64 \
  --test_ckpt_addr "${TEST_CKPT_ADDR}" --shot 0

Full-shot example (SN1):

python main_logofuse.py \
  --model ULIP_PointBERT --method logofuse --evaluate_3d \
  --dataset_name ShapeNetCore54 --dataset_split SN1 \
  --npoints 4096 --validate_dataset_prompt shapenet_64 \
  --test_ckpt_addr "${TEST_CKPT_ADDR}" --shot 999999

2. Synthetic-to-Real: ModelNet support -> ScanObjectNN test

Officially this track is used for SR1 and SR2.

The target/eval dataset remains ScanObjectNN15. Synthetic support is injected through a custom train .dat via --scanobject_train_dat.

The support .dat must already be prepared in a ScanObject-compatible label space.

Helper script:

bash tools/run_synth2real_sr12.sh /abs/path/to/modelnet_as_support_train.dat

Optional override for the ScanObject test dat:

SCANOBJECT_TEST_DAT=/abs/path/to/scanobject_test.dat \
  bash tools/run_synth2real_sr12.sh /abs/path/to/modelnet_as_support_train.dat

3. Real-to-Real: ScanObjectNN15

Zero-shot (SR1/SR2/SR3):

bash tools/run_zeroshot_sr123.sh

Full-shot (SR1/SR2/SR3):

bash tools/run_fullshot_sr123.sh

Negative-bank sweep (K_neg, full-shot):

bash tools/run_kneg_sweep_fullshot.sh 1 15

Verified Results

These numbers were re-verified with the cleaned mainline configuration.

ScanObjectNN15 full-shot

• SR1: AUROC 0.930279, FPR95 0.306383
• SR2: AUROC 0.919289, FPR95 0.369697
• SR3: AUROC 0.859724, FPR95 0.530387
• Macro Average: AUROC 0.903097, FPR95 0.402156

ShapeNetCore54 full-shot

• SN1: AUROC 0.913809, FPR95 0.464525
• SN2: AUROC 0.932384, FPR95 0.338373
• SN3: AUROC 0.963097, FPR95 0.248195
• Macro Average: AUROC 0.936430, FPR95 0.350364

Outputs

• logs: logs_*
• per-run summaries: results.tsv, summary.tsv
• feature cache: outputs/feature_cache_logofuse*

Repro Notes

• The helper scripts prefer ./.venv/bin/python when present.
• The public helper scripts force --rebuild_feature_cache by default. This avoids stale cached features from older runs or changed code paths from contaminating final metrics.
• If you run commands manually with --cache_features, prefer pairing it with --rebuild_feature_cache for any formal reproduction run.
• Full-shot support comes from the packaged train split, or from --scanobject_train_dat in the Synth-to-Real track.
• Zero-shot uses the same pipeline with --shot 0.
• The public scripts are aligned with the cleaned mainline configuration and do not rely on removed historical flags.