Paper Publication

Coming soon

Description

In Computer Vision, Convolutional Neural Networks (CNNs) have been widely used as a classic architecture. However, regarding image and video colorization, CNNs face challenges in capturing the relationships between pixels within a video frame due to the local nature of their kernels. This limitation often leads to inaccurate colorization results. In recent years, Vision Transformers (ViT) and their variants have emerged as promising solutions to address this issue.

In our research project, titled "Video Colorization using Swin Transformer," we introduce an end-to-end network designed explicitly for colorizing videos. Instead of utilizing the original ViT, we employ the Swin Transformer as the backbone for feature extraction. This novel approach represents the first utilization of the Swin Transformer backbone for video colorization tasks, yielding positive results.

By leveraging the power of the Attention mechanism, our proposed method effectively captures the relationships among all pixels within a video frame. This comprehensive understanding of pixel interactions enables a more accurate and efficient colorization process. To evaluate the performance of our approach, we conducted experiments using the DAVIS dataset, and the results demonstrate superior performance compared to state-of-the-art (SOTA) methods.

To further advance our research and perform inference on the "Video Colorization using Swin Transformer" model, we kindly request GPU resources from HuggingFace. The computational demands of processing video frames using the Swin Transformer can be substantial, and GPUs offer the necessary parallel processing capabilities to expedite the inference process. With the allocated GPU resources, we aim to validate and showcase the enhanced performance of our model on various video datasets, contributing to the advancement of video colorization techniques.

We sincerely appreciate your consideration and support in providing the requested GPU resources, which will significantly contribute to the success and impact of our research project.

Desired GPU type supported

We believe that the allocation of Tesla T4 would greatly facilitate our work and contribute to the success of our project.

Thanks for considering this request and hope to receive HuggingFace team's response soon.