Continuous Spatiotemporal Events Decoupling through Spike-based Bayesian Computation

Authors: Yajing Zheng, Jiyuan Zhang, Zhaofei Yu, Tiejun Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the constructed spiking network can effectively segment the motion contained in event streams.
Researcher Affiliation Academia 1 State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University 2 Institute for Artificial Intelligence, Peking University
Pseudocode Yes Algorithm 1 Spike-based Bayesian Computation for Event Motion Segmentation
Open Source Code No We will release the code and testing data after the paper is accepted.
Open Datasets Yes Fig. 4 shows the parameter initialization process in a sequence from the Extreme Event Dataset (EED) [26], which includes both camera self-motion and a moving object.
Dataset Splits No After initializing the parameters, we select a fixed number of events in chronological order, dividing all events into different packets {en}Ng n=1 as inputs to the network over time.
Hardware Specification No The primary focus is on CPU-based verification due to minimal graphical operations, ensuring fast processing as an online learning algorithm. Additionally, the speed on both GPU and CPU is comparable.
Software Dependencies No No specific software dependencies with version numbers are mentioned in the paper.
Experiment Setup Yes Parameter Initialization. The proposed spike-based Bayesian Computation framework and its corresponding event-based clustering framework are both locally convergent algorithms. ... Network Learning and Inference. After initializing the parameters, we select a fixed number of events in chronological order, dividing all events into different packets {en}Ng n=1 as inputs to the network over time. During online learning, we also split the n-th events packet {en} into different patches and feed them into the network. After optimizing the parameters θ for several epochs, we obtain the optimized motion parameters θ , and then input all events into the network to get the responsibilities P of all events belonging to different motion parameters.