AsymLoc: Asymmetric Point Supervision for Cross-Dataset Zero-Shot 3D Anomaly Localization

Cross-dataset zero-shot 3D anomaly localization aims to identify defective regions in previously unseen object categories without accessing target-domain data during training. Although multi-view vision–language approaches provide transferable semantic representations, their point-level anomaly responses may remain insufficiently discriminative, and improving point-level localization does not necessarily improve instance-level discrimination. We propose AsymLoc, an asymmetric point supervision framework built upon multi-view projection with prompted Contrastive Language–Image Pre-training (MVP-PCLIP). The framework emphasizes anomalous points through positive supervision and uses selective hard-normal regularization as an auxiliary constraint on high-scoring normal regions. A decoupled instance-level score fusion module combines view-level semantic evidence with point-level anomaly responses while preventing sample-level gradients from rewriting the localization map. Under the MVTec 3D-AD→Real3D-AD protocol, AsymLoc improves point-level AUROC, F1 score, and average precision by 2.8, 7.6, and 7.0 points over a reproduced MVP-PCLIP baseline, and instance-level AUROC by 8.4 points. On the reverse Real3D-AD→MVTec 3D-AD direction, point-level metrics remain close to the matched baseline while instance-level AUROC rises by 3.7 points. Positive supervision accounts for most of the localization improvement; hard-normal regularization is an auxiliary constraint; decoupled fusion mainly affects instance ranking.