Roto-translated Local Coordinate Frames For Interacting Dynamical Systems

Authors: Miltiadis Kofinas, Naveen Nagaraja, Efstratios Gavves

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments in traffic scenes, 3D motion capture, and colliding particles demonstrate that the proposed approach comfortably outperforms the recent state-of-the-art. 5 Experiments We evaluate the proposed method, Lo CS, on 2D and 3D geometric graph dynamical systems from the literature. In 2D, we evaluate on a synthetic physics simulation dataset proposed by d NRI [19] and on traffic trajectory forecasting [4]. In 3D, we evaluate on an 3D-extended version of the charged particles [27] and on a motion capture dataset [10]. We compare with NRI [27], d NRI [19], and the very recent EGNN [41].
Researcher Affiliation Collaboration Miltiadis Kofinas University of Amsterdam m.kofinas@uva.nl Naveen Shankar Nagaraja BMW Group Naveen-Shankar.Nagaraja@bmw.de Eftratios Gavves University of Amsterdam egavves@uva.nl
Pseudocode No The paper describes methods in text and equations, but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Our code, data, and models will be available online1. 1https://github.com/mkofinas/locs
Open Datasets Yes On the synthetic 2D physics simulation we use the same experimental settings as in [19]. In the charged particles datasets the particles interact with one another via electrostatic forces. We extend the dataset by [27] from 2 to 3 dimensions... The in D dataset [4] is a real-world 2D traffic trajectory forecasting dataset... Last, we experiment with the CMU motion capture database [10] with 3D data...
Dataset Splits Yes We generate 30,000 scenes for training, 5,000 for validation and 5,000 for testing. The dataset contains 36 recordings; we split them in 19/7/10 for training, validation and testing.
Hardware Specification No The paper does not specify any particular hardware used for experiments (e.g., GPU/CPU models, memory details).
Software Dependencies No The paper mentions software components like MLPs, LSTMs, and Gumbel-Softmax, and refers to code of other methods, but does not provide specific version numbers for any software dependencies used for their own implementation.
Experiment Setup No The paper states 'The full implementations details are in appendix B.3.' but does not provide specific hyperparameter values or detailed training configurations within the main text.