Back to 3D: Few-Shot 3D Keypoint Detection with Back-Projected 2D Features

TL;DR

This work introduces B2-3D, a learning-free yet powerful approach for few-shot 3D keypoint detection by back-projecting rich 2D foundation-model features onto 3D surfaces from multiple views. A key contribution is a global keypoint optimization that aligns per-keypoint features while preserving their geometric relationships via pairwise geodesic distances, enabling robust handling of symmetries. Empirically, the method achieves state-of-the-art performance on KeypointNet (large IoU gains across thresholds) and strong results in ShapeNet Part segmentation transfer, illustrating strong generalization. The study also analyzes the properties of back-projected features, showing stability to rendering choices, semantic richness, and sensitivity to local geometry, thereby establishing back-projected 2D features as a versatile prior for various 3D shape analysis tasks.

Abstract

With the immense growth of dataset sizes and computing resources in recent years, so-called foundation models have become popular in NLP and vision tasks. In this work, we propose to explore foundation models for the task of keypoint detection on 3D shapes. A unique characteristic of keypoint detection is that it requires semantic and geometric awareness while demanding high localization accuracy. To address this problem, we propose, first, to back-project features from large pre-trained 2D vision models onto 3D shapes and employ them for this task. We show that we obtain robust 3D features that contain rich semantic information and analyze multiple candidate features stemming from different 2D foundation models. Second, we employ a keypoint candidate optimization module which aims to match the average observed distribution of keypoints on the shape and is guided by the back-projected features. The resulting approach achieves a new state of the art for few-shot keypoint detection on the KeyPointNet dataset, almost doubling the performance of the previous best methods.

Figures (14)

• Figure 1: Qualitative results of our proposed method B2-3D for few-shot keypoint detection using back-projected features (red) with ground truth keypoint annotations (green).
• Figure 2: After processing the rendered views of an object with a large pre-trained vision encoder (e.g., DINO), we back-project the features onto the 3D shape and aggregate the information from all views (a) to obtain rich semantic 3D features (b).
• Figure 3: Feature stability analysis measuring the mean cosine similarity (with standard deviation in light blue) of extracted point features when applying modifications to the rendering process.
• Figure 4: Increase in feature quality and distinctiveness with increasing number of rendering viewpoints. Visualization using a PCA, as described in Sec. \ref{['sec:semantic-analysis']}.
• Figure 5: Back-projected ViT features on a shape can be visualized after performing a PCA to just three values per vertex. The extracted features contain rich semantic information and clearly assign different values to different semantic parts of the object.