Learning with 3D rotations, a hitchhiker’s guide to SO(3)
Authors: Andreas René Geist, Jonas Frey, Mikel Zhobro, Anna Levina, Georg Martius
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we empirically assess the earlier discussion and support the recommendations with various experiments on rotation estimation and feature prediction. Experiment 1 (rotation estimation): Rotation from point clouds, Experiment 2.1 (rotation estimation): Cube rotation from images, Experiment 3 (rotation estimation): 6D object pose estimation from RGB-D images, Experiment 4 (feature prediction): SO(3) as input to Fourier series. |
| Researcher Affiliation | Academia | 1Max Planck Institute for Intelligent Systems, Autonomous learning group, Germany 2University of Tübingen, Distributed intelligence lab, Germany 3Swiss Federal Institute of Technology (ETH Zurich), Robotic systems lab, Switzerland 4University of Tübingen, Self-organization of neuronal networks group, Germany. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The project code is available at: github.com/martius-lab/hitchhiking-rotations |
| Open Datasets | Yes | Point clouds are extracted from a set of 726 airplane CAD models... using 'Sci Py' (Virtanen et al., 2020). We use the implementation provided of Wang et al. (2019) and test the performance of the network on the YCB-Video Dataset... One experiment found in Zhou et al. (2019); Levinson et al. (2020); Brégier (2021); Pepe et al. (2022) is Inverse Kinematics with CMU Mo Cap data (see Sec. E.5). |
| Dataset Splits | Yes | We keep the number of points fixed to 800, 200, and 1000 for the train, the validation, and the test data set, respectively. |
| Hardware Specification | Yes | All numbers are reported using an Nvidia RTX3060. |
| Software Dependencies | Yes | Point clouds are extracted from a set of 726 airplane CAD models... using 'Sci Py' (Virtanen et al., 2020). Adam (standard settings in Py Torch with a starting learning rate of 0.001). |
| Experiment Setup | Yes | We incorporate early stopping in our training process with a patience of 10 epochs and a maximum of 100 epochs... In all experiments, we use the Adam optimizer and train for up to 1000 epochs. ... For the orientation prediction task, we use a learning rate of 0.001 and a batch size of 32. For the feature prediction/image generation task, we use a learning rate of 0.01 and a batch size of 128. |