E-Motion: Future Motion Simulation via Event Sequence Diffusion

Authors: Song Wu, Zhiyu Zhu, Junhui Hou, GUANGMING Shi, Jinjian Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive testing and validation, we demonstrate the effectiveness of our method in various complex scenarios, showcasing its potential to revolutionize motion flow prediction in computer vision applications such as autonomous vehicle guidance, robotic navigation, and interactive media. Our findings suggest a promising direction for future research in enhancing the interpretative power and predictive accuracy of computer vision systems.
Researcher Affiliation Academia Song Wu 1, Zhiyu Zhu 2, Junhui Hou 2, Guangming Shi 1 , Jinjian Wu 1 1 Xidian University, 2 City University of Hong Kong
Pseudocode Yes Algorithm 1 Motion Alignment Process
Open Source Code Yes The source code is publicly available at https://github.com/p4r4mount/E-Motion.
Open Datasets Yes In our study, we utilize two large-scale event datasets, i.e., Vis Event [55] and Event VOT dataset [56].
Dataset Splits Yes The dataset is meticulously annotated and divided into training (841 videos), validation (18 videos), and testing (282 videos) subsets, ensuring a comprehensive framework for robust algorithm testing and benchmarking.
Hardware Specification Yes All experiments are conducted on machines with 8 Ge Force RTX 3090 GPUs, Intel(R) Core(TM) i7-10700 CPU of 2.90GHz, and 64-GB RAM.
Software Dependencies No The paper mentions optimizers like ADAM and PPO, but it does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes In the pre-training stage, we employed the ADAM optimizer with the exponential decay rates β1 = 0.9 and β2 = 0.999. The total training process was 20000 iterations for both kinds of noise experiments. We initialized the learning rate as 1e-5. We set the batch size to 128 (with 8 gradient accumulation steps). For the alignment process... The updating episode of the reinforcement learning process is set at 100 optimization steps... We also employed the ADAM optimizer with the exponential decay rates β1 = 0.9 and β2 = 0.999. We initialized the learning rate as 2e 6 and set the batch size to 16(with 2 gradient accumulation steps) in all experiments.