FFT-XMem: Enhancing the Boundary Identification of Video Object Segmentation by Fast Fourier Transform Based XMem Network

As for video object segmentation, the recognition of various object and the corresponding position is essential to high-performance object segmentation of streaming video data. Meanwhile, semi-supervised video object segmentation aims to improve data preparation and annotation cost, and developing deep neural networks to process streaming images in spectral domain, which applies filter kernels in the spectral domain to recognize the boundaries of an image, is also an effective way to enhance the performance of video object segmentation. Thus, this paper purpose a simple yet effective concept for enhancing the boundary Identification among the segmentation objects, and the structure of the proposed method consists of Fast Fourier Transform and XMem for effectively recognizing the object boundary based on the features extracted from the spectral domain. The proposed module tries to provide all-around features of the object boundary, so this work applies the existing state-of-the-art network as a baseline for integrating the advantage of Fourier module. Through designing a highresolution object boundary extraction module, the purposed method can decrease the loading of computation hardware. This work trained the model on one RTX 4090 with two datasets, and the experimental results shows a significant boost in performance compared to other state-of-the-art models. Further, this method also demonstrates the scalability in different models and evidently proves the capability of a compact model in enhancing video object segmentation by low-level pixel matching and object boundary recognition.