Code and Dataset for: Geometry-Induced 3D Defect Detection for Automotive Connector Pins Using Neural Radiance Fields

This repository contains the code, processed dataset, and supporting materials for a connector-pin geometric anomaly detection framework. The main pipeline extracts slot-level position and direction tolerance features from reconstructed connector point clouds, then trains and evaluates a regularized LightGBM detector for bent or displaced pin detection. The dataset is provided as processed feature tables, including training features, eval-only test labels, and blind test features. The test set contains 47 connector samples and 940 canonical slots, including 900 present-pin slots for geometric defect evaluation and 40 missing-pin slots. The main directory structure is organized as follows: `programs/tolerance_frontend/`: scripts for tolerance evidence extraction, slot matching, feature-bank construction, and metric evaluation. - `programs/modeling/`: scripts for grouped cross-validation, model selection, threshold selection, and frozen test evaluation. - `configs/`: selected model configuration, tolerance-band configuration, feature catalog, and dataset manifest. - `data_tables/`: processed CSV feature tables, including `train_labeled_features.csv`, `test_labeled_features_eval_only.csv`, and `test_blind_features.csv`. - `evidence/`: final evaluation reports, error analysis files, and ablation results. - `docs/`: detailed algorithm notes and reproducibility documentation. - `FINAL/` and `picture/`: manuscript source files and figure assets used in the paper. To use the code, install the required Python packages such as `numpy`, `pandas`, `scikit-learn`, `scipy`, `open3d`, and optionally `lightgbm`. The processed feature tables in `data_tables/` can be used directly for model training and evaluation. Model selection is performed with the grouped out-of-fold script in `programs/modeling/`, and the frozen evaluation is performed using the selected configuration in `configs/latest_selected_config.json`. When adapting the code to a new directory layout, update the input paths in the modeling utilities accordingly. The test labels are eval-only and should not be used for feature selection, threshold tuning, or model selection.