Pith. sign in

REVIEW 2 cited by

CLIP goes 3D: Leveraging Prompt Tuning for Language Grounded 3D Recognition

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2303.11313 v3 pith:66GHFDGG submitted 2023-03-20 cs.CV

CLIP goes 3D: Leveraging Prompt Tuning for Language Grounded 3D Recognition

classification cs.CV
keywords clipcg3dfeaturesimagescapabilitiesencoderlanguagenatural
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-Language models like CLIP have been widely adopted for various tasks due to their impressive zero-shot capabilities. However, CLIP is not suitable for extracting 3D geometric features as it was trained on only images and text by natural language supervision. We work on addressing this limitation and propose a new framework termed CG3D (CLIP Goes 3D) where a 3D encoder is learned to exhibit zero-shot capabilities. CG3D is trained using triplets of pointclouds, corresponding rendered 2D images, and texts using natural language supervision. To align the features in a multimodal embedding space, we utilize contrastive loss on 3D features obtained from the 3D encoder, as well as visual and text features extracted from CLIP. We note that the natural images used to train CLIP and the rendered 2D images in CG3D have a distribution shift. Attempting to train the visual and text encoder to account for this shift results in catastrophic forgetting and a notable decrease in performance. To solve this, we employ prompt tuning and introduce trainable parameters in the input space to shift CLIP towards the 3D pre-training dataset utilized in CG3D. We extensively test our pre-trained CG3D framework and demonstrate its impressive capabilities in zero-shot, open scene understanding, and retrieval tasks. Further, it also serves as strong starting weights for fine-tuning in downstream 3D recognition tasks.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. PASR: Pose-Aware 3D Shape Retrieval from Occluded Single Views

    cs.CV 2026-04 unverdicted novelty 5.0

    PASR performs pose-aware analysis-by-synthesis by aligning 3D projections with DINOv3 patch features, outperforming prior methods on clean and occluded retrieval while also handling pose estimation and classification.

  2. 3DAlign-DAER: Dynamic Attention Policy and Efficient Retrieval Strategy for Fine-grained 3D-Text Alignment at Scale

    cs.CV 2025-11 unverdicted novelty 5.0

    3DAlign-DAER improves fine-grained 3D-text alignment at scale via dynamic attention policy with Monte Carlo calibration and efficient retrieval, supported by a new Align3D-2M dataset of 2M pairs.