Replication package for "A Controlled Study of Multi-Scale and Occlusion-Aware Modules in Direct 6D Pose Regression"

This replication package accompanies the manuscript "A Controlled Study ofMulti-Scale and Occlusion-Aware Modules in Direct 6D Pose Regression"(Frontiers in Mechanical Engineering), containing the data splits, sourcecode, per-run training reports, per-instance test predictions, anddocumentation needed to reproduce or re-evaluate every result reported inthe paper and in the response to reviewers. Contents:- data/splits/ per-seed train/val/test split files for LM and LM-O- scripts/ training, evaluation and diagnostic scripts, plus the synthetic-occlusion visibility-mask generator- src/ the model package and original training/evaluation code- results/ report.json for all 18 re-run configurations (3 module configurations x 2 datasets x 3 seeds), occlusion- stratified means, model cost measurements, root-cause diagnostics, and per-instance test predictions for the LM-O (1,517 instances) and LM (3,000 instances) test sets, including predicted and ground-truth poses for exact re-evaluation with the official BOP toolkit- results/demo_60ep/ and results/demo_150ep/ diagnostic demo runs under the corrected data protocol (checkpoint, report, log, diagnostics)- figures/ all figures used in the manuscript and the response- README.md full reproduction instructions (datasets, protocol, training, evaluation)- requirements_redo.txt pinned dependencies Key results reproduced from this package: under the original data protocolall configurations fail catastrophically on the real occlusion test set ofLM-O (mean ADD about 57 cm), while validation ADD is misleadingly small(< 1 cm); a minimal diagnostic experiment that removes the two root causes(pose-diverse PBR training data and a scale-preserving fixed-window crop)improves test ADD from 57.3 to 22.5 cm. The failure is attributable to thedata protocol, not to the architecture. The BOP benchmark datasets are not redistributed here (see README fordownload instructions); the synthetic-occlusion training set is regenerablefrom the provided scripts.