3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction
Authors: Jongmin Lee, Minsu Cho
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method achieves state-of-the-art performance on standard single object pose estimation benchmarks, including Model Net10-SO(3) and PASCAL3D+, demonstrating high sampling efficiency and strong generalization to unseen 3D rotations. |
| Researcher Affiliation | Collaboration | Pohang University of Science and Technology (POSTECH), South Korea {ljm1121, mscho}@postech.ac.kr *The current affiliation of Jongmin Lee is with LG AI Research. Contact: jongminlee@lgresearch.ai. |
| Pseudocode | No | The paper does not contain a pseudocode block or a clearly labeled algorithm section. |
| Open Source Code | No | We will also make our code publicly available after acceptance. |
| Open Datasets | Yes | Model Net10-SO(3) [42] is a common dataset for estimating a 3D rotation from a single image. ... PASCAL3D+ [68] is a widely-used benchmark for evaluating pose estimation in images captured in real-world settings. |
| Dataset Splits | Yes | We use a maximum frequency level of L = 6, resulting in a total size of M = 455, computed as P6 l=0(2l + 1) (2l + 1). ...At inference, we employ a recursive level 5 of SO(3) HEALPix grid with 2.36 million points, achieving a precision of 1.875 , as in [28, 35]. |
| Hardware Specification | Yes | We use a machine with an Intel i7-8700 CPU and an NVIDIA Ge Force RTX 3090 GPU. |
| Software Dependencies | No | We utilize the e3nn library [22] for S2 and SO(3) convolutions for efficient handling of both Fourier and inverse Fourier transforms, healpy [24, 81] for HEALPix grid generation, and Py Torch [53] for model implementation. |
| Experiment Setup | Yes | With a batch size of 64, our network is trained for 50 epochs on Model Net10-SO(3) taking 25 hours, and for 80 epochs on PASCAL3D+ taking 28 hours. We start with an initial learning rate of 0.1, which decays by a factor of 0.1 every 30 epochs. We use the SGD optimizer with Nesterov momentum set at 0.9. |