PRS-Net: 3D Reflective Symmetry Detection

• PRS-Net is a 3D convolutional network that regresses explicit symmetry parameters from voxelized models to detect global planar reflective symmetries and generalized cylinders.
• It employs a convolutional backbone with dual MLP branches to predict reflection planes and rotation axes, achieving high precision and rapid inference.
• Its unsupervised training using a combined symmetry distance and regularization loss makes it robust to noise, partial data, and variations in geometric datasets.

PRS-Net refers to "Planar Reflective Symmetry Detection Net for 3D Models," a 3D convolutional neural network architecture designed for unsupervised discovery of global planar reflective symmetries and generalized cylinders in 3D shapes. The method conceptualizes symmetry detection as a regression problem over explicit symmetry parameters, rather than relying on traditional spatial sampling strategies or annotation-driven learning. The network processes voxelized shape representations and delivers high-precision, high-speed symmetry characterization, demonstrating robustness to noise and partial data across large geometric datasets (Gao et al., 2019).

1. Network Architecture and Input Representation

PRS-Net inputs a 32×32×32 occupancy voxel grid $V \in \{0,1\}^{32 \times 32 \times 32}$ produced by normalizing a mesh or point cloud to the unit cube and voxelizing its occupancy. The architecture consists of the following components:

• Convolutional Backbone: Five consecutive 3D convolutional blocks, each comprising a $3\times3\times3$ convolution (stride 1, padding 1) with increasing feature channels ($C_1=16$, $C_2=32$, $C_3=64$, $C_4=128$, $C_5=256$), followed by LeakyReLU activations (slope 0.2) and $2\times2\times2$ max pooling (stride 2).
• Feature Aggregation: After five rounds of pooling, the feature volume collapses to size $1 \times 1 \times 1 \times 256$, forming a 256-dimensional vector.
• Prediction Heads: Two parallel fully connected branches (MLPs):
  • One branch predicts three reflection planes $P_i = (a_i, b_i, c_i, d_i)$, $3\times3\times3$0;
  • The other predicts three rotation axes as quaternions $3\times3\times3$1, $3\times3\times3$2.
• Normalization: Predicted quaternions are $3\times3\times3$3-normalized post-forward-pass to ensure valid rotation representations.

2. Symmetry Parameterization and Initialization

• Reflection Planes: Each output plane $3\times3\times3$4 is parameterized as $3\times3\times3$5, with $3\times3\times3$6 the (unconstrained) normal during training, normalized as needed for loss calculations, and $3\times3\times3$7 the signed offset.
• Rotation Symmetries: Each rotation axis is predicted as a unit quaternion, which, after training, is decomposed to axis and angle as $3\times3\times3$8, $3\times3\times3$9.
• Initialization: Plane normals and quaternions are seeded as the three principal axes and corresponding $C_1=16$0 rotations: $C_1=16$1, $C_1=16$2, $C_1=16$3.

3. Loss Formulation and Training

PRS-Net is trained fully unsupervised, with no symmetry annotations, via a two-term loss:

• Symmetry Distance Loss ($C_1=16$4): For each predicted symmetry (plane or axis), $C_1=16$5 points $C_1=16$6 are uniformly sampled from the surface. Each $C_1=16$7 is reflected or rotated using the predicted parameters, and the Euclidean distance $C_1=16$8 to the nearest surface point is computed. This is summed across three planes and three axes:

$C_1=16$9

• Regularization Loss ($C_2=32$0): Prevents degenerate outputs (e.g., repeated planes/axes) using orthogonality of predicted normal and axis matrices $C_2=32$1, $C_2=32$2:

$C_2=32$3

• Full Objective: $C_2=32$4, where $C_2=32$5 yields empirically best non-redundant symmetries.

In training, augmentation with 4000 random rotations per category and pre-processing ensures invariance and coverage. Training proceeds on large datasets (e.g., ShapeNet, ABC, Thingi10K), using Adam with learning rate $C_2=32$6, batch size $C_2=32$7, and observed convergence in approximately $C_2=32$8 steps (reducible to $C_2=32$9 with pre-training on Kazhdan et al.'s symmetry annotations).

4. Postprocessing and Result Filtering

Since the network always predicts three planes/axes, postprocessing eliminates redundant or low-quality symmetries:

• Duplicate Removal: For plane pairs with dihedral angle $C_3=64$0, the one with higher average surface-mapping error $C_3=64$1 is discarded.
• Thresholding: Planes or axes with mean $C_3=64$2 (normalized units) are removed.
• Generalized Cylinders: For rotation axes, PRS-Net checks whether the predicted axis corresponds to a generalized cylinder by evaluating sampled rotations (each $C_3=64$3), keeping axes whose induced symmetry error remains below the threshold.

5. Experimental Results and Performance Analysis

In quantitative benchmarks, PRS-Net outperforms state-of-the-art sampling-based and analytic symmetry detectors:

• ShapeNet (1,000 models): Achieves Ground-Truth Error (GTE) $C_3=64$4 versus e.g., Kazhdan et al. $C_3=64$5, Martinet et al. $C_3=64$6, and Mitra et al. $C_3=64$7.
• Symmetry Distance Error (SDE): $C_3=64$8, better or comparable to analytic approaches (Kazhdan $C_3=64$9).
• Generalization: On ABC (CAD) models, SDE $C_4=128$0 versus next best $C_4=128$1–$C_4=128$2; On Thingi10K, $C_4=128$3.
• Noisy/Partial Inputs: Robustness demonstrated by low SDE with missing data or vertex-normal Gaussian noise: SDE increases only slightly and remains lowest among tested methods.

Ablation studies reveal the $C_4=128$4 voxel grid provides optimal trade-off; higher or lower resolutions impair performance due to over- or under-fitting. The regularization weight $C_4=128$5 minimizes duplication without excessively enforcing orthogonality.

Inference Speed: PRS-Net processes a model in $C_4=128$6 ms on NVIDIA TITAN X GPU, two orders of magnitude faster than PRST (3.4 s) or Kazhdan (0.51 s).

6. Limitations and Future Directions

PRS-Net predicts a fixed, maximum number (three) of symmetry planes and axes per shape, which precludes adaptively inferring the actual number present. The method does not address intrinsic or higher-order rotational symmetries beyond global planar and cylindrical structures. Extension to predicting a variable number of symmetries, or to non-reflective, intrinsic, and higher-order symmetry forms, is noted as an open research avenue (e.g., via reinforcement learning head selection).

A plausible implication is that broad adoption in geometric deep learning pipelines may depend on overcoming these architectural constraints and integrating with more general shape-understanding frameworks.

7. Context and Significance in Geometry Processing

PRS-Net establishes a paradigm shift for geometry processing by reframing symmetry analysis as an unsupervised regression task over geometric parameters via 3D-CNNs, eschewing the computational cost and limitations of sampling-based or annotation-driven pipelines. Its real-time inference and robustness have potential impact across tasks such as mesh segmentation, shape matching, 3D object completion, and structural analysis. The avoidance of reliance on ground-truth symmetry annotation enables broader applicability to diverse datasets, including scanned, incomplete, and noisy representations (Gao et al., 2019).