Enhanced Back-Projection of Vision Features for 3D Symmetry Detection

We propose two algorithms for 3D symmetry detection based on enhanced back-projection of vision features extracted from foundation vision models such as DINOv2. Our method enhances back-projection by rendering multiple views of 3D objects, extracting features, and projecting them onto the geometry with two key improvements—Fibonacci view sampling and view rotations—that increase robustness and accuracy. Using these features, we detect symmetry planes and axes through two dedicated algorithms. Experiments on ShapeNet show that our plane detection approach outperforms both traditional geometric and learning-based methods by a wide margin. The method is also efficient, running in seconds on a single 8GB GPU, making it practical for large-scale or real-world applications. Overall, our results demonstrate that enhanced back-projection of vision features offers a simple yet effective framework for solving fundamental 3D geometric problems such as symmetry detection.

Diagram showing the two key enhancements to back-projection: Fibonacci view sampling and view rotations. These improvements increase the robustness and accuracy of the projected features, leading to better symmetry detection results.

This produces different feature representations: Raw-Mesh (RM), Point Cloud (PC), Feature-Mesh Sampling (FM), where FM performs the same sampling for geometry and features after RM.

The plane detection algorithm follows these steps:

1. Group each point into a trio of similar features
2. Compute 4 possible candidate planes
3. Filter planes that pass near the middle of the object, as these represent global symmetries
4. Check resulting planes using Chamfer distance and output the best computed planes

The axis detection algorithm follows these steps:

1. Group each point into an L-sized group of similar features
2. Use RANSAC to produce circles for each group
3. Cluster resulting circles from all groups and keep the largest cluster
4. Compute a mean circle and output the axis defined by the mean circle's center and normal vector

Both algorithms use 114 views with DINOv2 small. Planes use FM10K with Fibonacci sampling and 4 rotations. Axes use RM with Fibonacci sampling and 4 rotations.

Plane Detection Results

FM10K (left) and RM (right), where the right image is for visual reference only to show the plane detected using features from the left image. Red outlines indicate detected symmetry planes.

Axis Detection Results

Qualitative examples using raw mesh representation. Blue outlines indicate detected rotational symmetry axes.