CEDFlow: Latent Contour Enhancement for Dark Optical Flow Estimation

Authors: Fengyuan Zuo, Zhaolin Xiao, Haiyan Jin, Haonan Su

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on the FCDN and VBOF datasets demonstrate that CEDFlow outperforms state-of-the-art methods in terms of the EPE index and produces more accurate and robust flow estimation.
Researcher Affiliation Academia Fengyuan Zuo1, Zhaolin Xiao1, 2*, Haiyan Jin1, 2, Haonan Su1, 2 1Xi an University of Technology, China, 710048 2Shaanxi Key Laboratory for Network Computing and Security Technology, China, 710048 xiaozhaolin@xaut.edu.cn, jinhaiyan@xaut.edu.cn
Pseudocode No The paper describes algorithms and shows diagrams but does not include pseudocode blocks or algorithms explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes Our code is available at: https://github.com/xautstuzfy.
Open Datasets Yes Experimental results on the FCDN and VBOF datasets demonstrate that CEDFlow outperforms state-of-the-art methods in terms of the EPE index and produces more accurate and robust flow estimation. Zheng et al. propose a synthetic optical flow benchmark by adding dark image noise to the Fly Chairs dataset, called Flying Chairs Dark&Noise (FCDN) dataset (Zheng, Zhang, and Lu 2020). They also introduce the Various Brightness Optical Flow (VBOF) dataset, which includes multiple exposure levels and optical flow pseudo labels (Zhang, Zheng, and Lu 2021).
Dataset Splits No The paper mentions training on FCDN and Mix (FCDN + VBOF) datasets and evaluating on FCDN and VBOF, but does not explicitly provide details about training/validation/test splits, percentages, or validation sets.
Hardware Specification No The paper discusses computational costs, runtime, and memory requirements of the model but does not specify any particular CPU, GPU, or other hardware models used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions like PyTorch, TensorFlow, or CUDA).
Experiment Setup Yes The suggested parameter settings for the FCDN and VBOF datasets are r = 3.0, resulting in a Gaussian kernel size of 7 × 7, and σ1 = 3, σ2 = 9, σ3 = 27 respectively. In our experiments, we set γ = 0.9 cooperating with many flow prediction iterations (K = 12), enabling a better coarse-to-fine flow updating.