Self-Supervised Learning of Event-Based Optical Flow with Spiking Neural Networks

Authors: Jesse Hagenaars, Federico Paredes-Valles, Guido de Croon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments with various types of recurrent ANNs and SNNs using the proposed pipeline. We validate our proposals through extensive quantitative and qualitative evaluations on multiple datasets.
Researcher Affiliation Academia Micro Air Vehicle Laboratory Delft University of Technology, The Netherlands
Pseudocode No The paper describes network architectures and processes but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes The project s code and additional material can be found at https://mavlab.tudelft.nl/event_flow/.
Open Datasets Yes train our networks on the indoor forward-facing sequences from the UZH-FPV Drone Racing Dataset [12], which is characterized by a much wider distribution of optical flow vectors than the datasets that we use for evaluation, i.e., MVSEC [54], High Quality Frames (HQF) [44], and the Event-Camera Dataset (ECD) [31].
Dataset Splits No The paper does not explicitly define training/validation/test splits with percentages or counts for its own data. It uses different datasets for training and evaluation, but a dedicated validation split from the training data is not specified.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running experiments.
Software Dependencies No Our framework is implemented in Py Torch. The paper mentions PyTorch but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We use the Adam optimizer [24] and a learning rate of 0.0002, and train with a batch size of 8 for 100 epochs. We clip gradients based on a global norm of 100. We fix the number of events for each input partition to N = 1k, while we use 10k events for each training event partition. Lastly, we empirically set the scaling weight for Lsmooth to λ = 0.001.