Exportar Publicação

Ramna Maqsood, Nunes, P., Soares, L. D. & Conti, C. (2025). Efficient Frequency-Aware Multiscale Vision Transformer for Event-to-Video Reconstruction. 2025 33rd European Signal Processing Conference (EUSIPCO).

R. Maqsood et al.,  "Efficient Frequency-Aware Multiscale Vision Transformer for Event-to-Video Reconstruction", in 2025 33rd European Signal Processing Conf. (EUSIPCO), Palermo, Italy, 2025

@misc{maqsood2025_1788417304201,
	author = "Ramna Maqsood and Nunes, P. and Soares, L. D. and Conti, C.",
	title = "Efficient Frequency-Aware Multiscale Vision Transformer for Event-to-Video Reconstruction",
	year = "2025",
	url = "https://ieeexplore.ieee.org/document/11226686"
}

TY  - CPAPER
TI  - Efficient Frequency-Aware Multiscale Vision Transformer for Event-to-Video Reconstruction
T2  - 2025 33rd European Signal Processing Conference (EUSIPCO)
AU  - Ramna Maqsood
AU  - Nunes, P.
AU  - Soares, L. D.
AU  - Conti, C.
PY  - 2025
CY  - Palermo, Italy
UR  - https://ieeexplore.ieee.org/document/11226686
AB  - Event-to-video (E2V) reconstruction is a critical task in event-based vision, benefiting from the advantages of event cameras, such as high dynamic range and low latency. However, existing deep learning reconstruction methods often prioritize temporal consistency and over-emphasize low-frequency features, leading to blur artifacts and loss of fine details. To overcome these limitations, we propose a novel frequency-aware multiscale vision transformer model for E2V reconstruction (MSViT-E2V). Our model employs wavelet-based decomposition to extract features at multiple scales, preserving fine-grained details through multilevel wavelet-based downsampling blocks, followed by transformer blocks for multiscale feature aggregation and long-range dependency modeling. Extensive experiments on various event datasets demonstrate that our model not only minimizes artifacts and preserves fine details but also reduces computational costs by up to 50% compared to the transformer-based model ET-Net.
ER  -