Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation

Authors: Xin Hu, Xiaole Tang, Ruixuan Yu, Jian Sun

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our model achieves state-of-the-art performance on multiple surface reconstruction datasets, and also exhibits better generalization to crossdataset shapes and robustness to arbitrary rotations.
Researcher Affiliation Collaboration Xin Hu1, Xiaole Tang1, Ruixuan Yu2, Jian Sun(B)1,3 1 Xi an Jiaotong University, Xi an, China 2 Shandong University, Weihai, China 3 Pazhou Laboratory (Huangpu), Guangzhou, China
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code will be available at https://github.com/math Xin112/PEIF.git.
Open Datasets Yes Datasets. We experiment on four datasets including Shape Net [51], ABC [52], Synthetic Rooms [19], MGN [53]. (1) Shape Net [51], as pre-processed by [7], contains watertight meshes of shapes in 13 classes. ... (2) ABC [52] has one million CAD models, mainly mechanical objects. We use the splits from [54] and select watertight meshes for experiments: 3599/883/98 shapes for training/validation/testing. (3) Synthetic Rooms [19] contains 5k synthetic room scenes composed of random walls, floors, and Shape Net objects. We adopt the same train/validation/test division in [19]. (4) MGN [53] is a real scanned dataset containing 5 clothing categories.
Dataset Splits Yes We use the splits from [54] and select watertight meshes for experiments: 3599/883/98 shapes for training/validation/testing.
Hardware Specification Yes We conducted all experiments on one NVIDIA RTX 4090 GPU.
Software Dependencies No We implement our PEIF in Pytorch [49] using Adam optimizer [50]. No specific version numbers for PyTorch or Adam are provided.
Experiment Setup Yes The learning rate is 8 × 10−4. For each query point, the size of the neighborhood is set as K = 32 for Shape Net [51] and ABC [52] datasets, K = 54 for Synthetic Rooms [19] dataset. We set β = 0.1 in the training loss and Nm = 4 for the memory bank. Please refer to the Appendix for details on the structures of involved MLPs, and the effect of different values of β. ... MLPs. γθs in Eqn. ( 7) consists of four 1 × 1 convolution layers with 6, 32, 64, and 128 hidden units. γθq and γθq in Eqn. (8) consists of 1 × 1 convolution layers with 3, 32, 64, and 128 hidden units while these for γθp are 3, 32, 64, and 128. For γθa in Eqn. ( 14), the unit numbers are 256, 512, 256, and 384. For γθd, the unit numbers are 256, 256, 256, and 256. All feature dimensions are 128. For the multi-head memory bank M, the number of the memory bank is set as NM = 4, with each memory bank containing 596 items.