TeDA: Boosting Vision-Lanuage Models for Zero-Shot 3D Object Retrieval via Testing-time Distribution Alignment

TL;DR

TeDA tackles the challenge of open-set 3D object retrieval by adapting a pretrained vision-language model (CLIP) at test time to unknown categories. It converts 3D objects to multi-view images, extracts CLIP features, and refines query embeddings through an unsupervised, self-boosting process that aligns query-target distributions via KL divergence, while incorporating textual cues from InternVL to enrich representations. The method achieves state-of-the-art results on four open-set 3DOR benchmarks in both training-required and training-free settings and extends to depth-map inputs, highlighting the practicality of test-time VLM adaptation for 3D tasks. Overall, TeDA demonstrates that training-free, cross-modal alignment at test time can yield substantial gains in 3D retrieval without large 3D datasets or fine-tuning, enabling scalable zero-shot performance in real-world scenarios.

Abstract

Learning discriminative 3D representations that generalize well to unknown testing categories is an emerging requirement for many real-world 3D applications. Existing well-established methods often struggle to attain this goal due to insufficient 3D training data from broader concepts. Meanwhile, pre-trained large vision-language models (e.g., CLIP) have shown remarkable zero-shot generalization capabilities. Yet, they are limited in extracting suitable 3D representations due to substantial gaps between their 2D training and 3D testing distributions. To address these challenges, we propose Testing-time Distribution Alignment (TeDA), a novel framework that adapts a pretrained 2D vision-language model CLIP for unknown 3D object retrieval at test time. To our knowledge, it is the first work that studies the test-time adaptation of a vision-language model for 3D feature learning. TeDA projects 3D objects into multi-view images, extracts features using CLIP, and refines 3D query embeddings with an iterative optimization strategy by confident query-target sample pairs in a self-boosting manner. Additionally, TeDA integrates textual descriptions generated by a multimodal language model (InternVL) to enhance 3D object understanding, leveraging CLIP's aligned feature space to fuse visual and textual cues. Extensive experiments on four open-set 3D object retrieval benchmarks demonstrate that TeDA greatly outperforms state-of-the-art methods, even those requiring extensive training. We also experimented with depth maps on Objaverse-LVIS, further validating its effectiveness. Code is available at https://github.com/wangzhichuan123/TeDA.

Figures (7)

• Figure 1: (a) Existing methods like Uni3D train a point cloud encoder by aligning visual and textual encoders from CLIP using large-scale data triplets, which requires substantial time and GPU resources. In contrast, (b) our method leverages rapid test-time adaptation to narrow the gap between query and target instances, without the need for training or model adjustments, preserving the inherent priors and generalization capabilities of CLIP. As shown in (c), our method achieves a significant improvement in retrieval performance.
• Figure 2: Overview of our TeDA. It can be decomposed into two stages, with the first stage representing the query and target streams of 3D objects with off-the-shelf pretraind VLMs (i.e., CLIP and InternVL) while the other one applying test-time distribution alignment for more discriminative 3D descriptors.
• Figure 3: The precision-recall curve comparisons on the four datasets, respectively.
• Figure 4: Impact of View Numbers on OS-MN40-core.
• Figure 5: Impact of Fusion Weight $\lambda$.