Domain Generalizable Adaptation of 3D Vision-Language Models via Regularized Fine-Tuning¶
🕒 Published (v1): 2026-06-16 20:31 UTC · Source: Arxiv · Venue: TMLR · link
Why this paper was selected
TMLR; regularized fine-tuning for domain-generalizable 3D vision-language model adaptation
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
ReFine3D is a regularized fine-tuning framework for 3D vision-language models (specifically ULIP/ULIP-2) that addresses domain generalization under limited labeled data. It combines layer-selective fine-tuning with augmentation consistency regularization, LLM-generated text diversity, point-rendered vision supervision, and test-time augmentation. On standard 3D domain generalization benchmarks it consistently outperforms the prior SOTA (PointPRC) across base-to-new, cross-dataset, corruption, and few-shot settings.
Problem¶
Adapting tri-modal 3D VLMs (point cloud + image + text) to downstream domains with scarce labels causes overfitting and catastrophic forgetting. Existing PEFT methods (e.g., PointPRC) naively port image-domain prompt tuning to 3D, overlook the structural differences of unordered point clouds, and discard the pre-trained CLIP image encoder during fine-tuning—losing rich visual priors. Simultaneously, the smaller capacity of 3D encoders (PointBERT: 22.8 M vs. CLIP ViT-B/16: 86 M) makes them behave differently from image encoders under limited-data PEFT.
Method¶
ReFine3D operates on a frozen CLIP text encoder and frozen CLIP ViT-B/16 image encoder, fine-tuning only the last layer of a 12-layer PointBERT encoder (layers 1–11 frozen).
Augmentation-based consistency regularization. \(K=16\) geometric augmentations (random \(z\)-axis rotation \([0°,360°]\), scale \([0.8,1.2]\), jitter \(\sigma=0.01\)) are applied to each input point cloud \(x\), yielding \(\{x_k\}_{k=1}^K\). The total loss is:
Text diversity regularization. For each class, 7 synonyms are retrieved from WordNet, then Qwen-2.5-7B-Instruct generates 7 descriptive sentences per synonym (prompt template: "A 3D point cloud of {synonym}"). Point cloud embeddings are aligned against the averaged diverse text embeddings rather than a single hand-crafted template.
Point-rendered vision supervision. 12 views per object are rendered via Blender/BlenderProc at 512Ă—512. Rendered image features from the frozen CLIP image encoder are pulled into a tri-modal contrastive alignment with \(z_P\) and \(z_T\), recovering the visual priors that prior methods discard.
Test-time augmentation. \(K_\text{test}=5\) augmented point cloud views and 7 text prompts per class are generated; the top-\(H=3\) embeddings ranked by maximum softmax probability are selected and aggregated by majority voting.
Key Contributions¶
- ReFine3D framework: layer-selective fine-tuning of 3D VLMs (only the last transformer layer is trainable) designed for domain-generalizable adaptation.
- Augmentation-based consistency regularization: averages CE loss across \(K=16\) augmented views to reduce memorization of specific instances.
- Text synonymization: WordNet + LLM-generated diverse descriptive sentences as class text embeddings, promoting semantic robustness.
- Point-rendered vision supervision: Blender-rendered images fed to the frozen CLIP image encoder re-introduce visual priors discarded by prior PEFT methods.
- Confidence-based test-time augmentation: top-\(H\) selection by max softmax probability with majority voting improves inference robustness.
Results¶
All improvements are reported over PointPRC (Sun et al., 2024a) as the prior SOTA, with ULIP-2 backbone unless noted.
- Base-to-new generalization (avg over 5 datasets): Base 84.71% vs. 83.18% (+1.53%), New 77.46% vs. 76.10% (+1.36%), HM 80.40% vs. 79.48% (+0.92%).
- Cross-dataset OOD transfer (avg over 5 target domains): 53.62% vs. 51.19% (+2.43%); source ShapeNetV2 → targets ModelNet40, ScanObjectNN variants, Omni3D.
- Corruption robustness (ModelNet-C, severity=2): +1.80% average over corruption types.
- Few-shot learning: +3.11% with 1-shot supervision; gains persist through 2, 4, 8, 16 shots.
- Per-dataset highlights (ULIP-2, base-to-new HM): ShapeNetCoreV2 90.26% vs. 88.34%; S-OBJ_BG 86.06% vs. 84.28%; S-PB_T50_RS 75.90% vs. 73.97%.
Limitations¶
- Incremental novelty: Authors explicitly acknowledge that none of the individual components are novel; contribution is their combination.
- Architecture dependency: Requires a tri-modal backbone (point cloud + image + text encoders simultaneously available); not applicable to unimodal or asymmetric 3D models.
- Rendering overhead: Blender-based point-to-image rendering adds preprocessing cost and requires offline rendering pipelines.
- Modest absolute gains: Improvements are 1–3 percentage points over PointPRC across most benchmarks.
- Narrow backbone evaluation: Experiments use PointBERT as the sole 3D encoder; generalization to other architectures (e.g., Point-MAE, Point-M2AE) is not tested.
- Limited corruption severity coverage: Corruption experiments use severity=2 only; behavior at higher severity levels is unreported.
Relevance to Vision-Language Models¶
ReFine3D directly extends the CLIP/ULIP paradigm—central to the VLM literature—into the 3D point cloud domain, showing that the tri-modal alignment backbone can be adapted with minimal parameter updates while preserving generalizable priors. The text diversity strategy (WordNet synonyms + LLM paraphrasing) is a domain-agnostic regularizer applicable to 2D VLM fine-tuning as well. The finding that retaining the frozen image encoder improves 3D adaptation reinforces a broader principle in VLM fine-tuning: pre-trained cross-modal supervision from non-target modalities should not be discarded. For researchers tracking VLMs, this paper demonstrates how CLIP-style models can be specialized for non-image sensing modalities under distribution shift, a direction increasingly relevant as VLMs are deployed in robotics and autonomous driving pipelines.