Rethinking High-speed Image Reconstruction Framework with Spike Camera

Authors: Kang Chen, Yajing Zheng, Tiejun Huang, Zhaofei Yu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct quantitative and qualitative analyses on the U-CALTECH and U-CIFAR datasets, with experiments showing that our proposed Spike CLIP achieves remarkable restoration performance under low-light conditions. Furthermore, the reconstructed images are well-aligned with the broader visual features needed for downstream tasks, ensuring more robust and versatile performance in challenging environments.
Researcher Affiliation Academia 1School of Computer Science, Peking University 2State Key Laboratory for Multimedia Information Processing, Peking University 3Institute for Artificial Intelligence, Peking University EMAIL, EMAIL
Pseudocode No The paper describes the framework and methods in detail using natural language and figures (e.g., Figure 2 for the overall framework, Figure 3 for HQ images generation pipeline), but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/chenkang455/Spike CLIP
Open Datasets Yes We conduct both quantitative and qualitative experiments on the UHSR dataset (Zhao et al. 2024a), which includes realworld spikes captured by the spike camera under low light from ultra-high-speed moving objects of the CALTECH (Fei-Fei, Fergus, and Perona 2004) and CIFAR (Krizhevsky, Hinton et al. 2009) datasets, referred to as U-CALTECH and U-CIFAR respectively.
Dataset Splits Yes Each dataset contains 5,000 spike-label training pairs and 1,000 test pairs, with 250 250 spatial resolution.
Hardware Specification Yes the overall experimental framework is built on the Py Torch platform and trained based on an NVIDIA 4090 GPU.
Software Dependencies No The paper states that "the overall experimental framework is built on the Py Torch platform" but does not specify the version number for PyTorch or any other software dependencies.
Experiment Setup Yes We employ the Adam optimizer with a learning rate of 4e-5 on the UCALTECH dataset and 1e-4 on the U-CIFAR dataset. The training process consists of 25 total epochs: 5 epochs for coarse reconstruction, 1 epoch for prompt optimization, and 19 epochs for fine reconstruction. We employ the voxelization technique to convert the spike stream from an initial length of 200 to 50.