Rotation-Adaptive Point Cloud Domain Generalization via Intricate Orientation Learning

TL;DR

This work tackles orientation-induced failures in cross-domain 3D point-cloud recognition by introducing a rotation-adaptive domain generalization framework. It alternates between intricate orientation mining, which identifies challenging rotations for each sample, and orientation-aware contrastive training that enforces rotation-consistent, discriminative representations via an EMA teacher. The approach integrates an orientation consistency loss and a margin separation loss, culminating in a final objective that improves both rotation robustness and category separation. Extensive experiments on PointDA and PointSegDA demonstrate state-of-the-art performance and heightened stability across 64 rotation scenarios, highlighting the method's practical impact for real-world 3D understanding under unpredictable orientations. Formally, the method operates over orientations $M_i\,\in\mathcal{O}\subseteq SO(3)$ and optimizes for generalizable features that remain reliable when tested on unseen, arbitrarily rotated domains.

Abstract

The vulnerability of 3D point cloud analysis to unpredictable rotations poses an open yet challenging problem: orientation-aware 3D domain generalization. Cross-domain robustness and adaptability of 3D representations are crucial but not easily achieved through rotation augmentation. Motivated by the inherent advantages of intricate orientations in enhancing generalizability, we propose an innovative rotation-adaptive domain generalization framework for 3D point cloud analysis. Our approach aims to alleviate orientational shifts by leveraging intricate samples in an iterative learning process. Specifically, we identify the most challenging rotation for each point cloud and construct an intricate orientation set by optimizing intricate orientations. Subsequently, we employ an orientation-aware contrastive learning framework that incorporates an orientation consistency loss and a margin separation loss, enabling effective learning of categorically discriminative and generalizable features with rotation consistency. Extensive experiments and ablations conducted on 3D cross-domain benchmarks firmly establish the state-of-the-art performance of our proposed approach in the context of orientation-aware 3D domain generalization.

Figures (5)

• Figure 1: t-SNE visualization of the feature spaces in the geometric encoder of DGCNN wang2019dynamic trained on ModelNet wu20153d before and after random rotation. (a) Random rotation perturbation leads to a chaotic cluster in the feature space, rendering it non-discriminative. (b) Even with rotation augmentation, the rotated shapes tend to be located in the low-density region of the feature space, which remains non-discriminative and less consistent.
• Figure 2: t-SNE visualization of the feature spaces trained with (a) random rotation augmentation and (b) our proposed intricate orientation training. For the sake of simplicity, we only visualize samples from three categories (table, chair, and sofa), which share part of distinguishing features (e.g., slim legs, plane seats, etc). Samples from ModelNet (the source domain) and ShapeNet (the target domain) are denoted by star ('$\star$') and plus ('$+$'), respectively. The gray dash line denotes the approximated decision boundaries.
• Figure 3: (a) Maximum Mean Discrepancy borgwardt2006integrating between the features of the ModelNet and ShapeNet subdomains in PointDA qin2019pointdan, learned with different orientation augmentations. (b) The mean consistency rate was computed using KL-divergence between multiple augmented variants and the mean accuracy on ShapeNet.
• Figure 4: Pipeline of our intricate orientation learning framework for point cloud classification, which alternatively optimizes between the intricate orientation mining (Left) and orientation-aware contrastive training (Right).
• Figure 5: t-SNE visualizations of the feature space (a) trained using only intricate orientation mining, and (b) further refined by orientation-aware contrastive training. Point features from ModelNet (the source domain) and ShapeNet (the target domain) are denoted by star ('$\star$') and plus ('$+$'), respectively.