Realtime Image Transformation Using Depth Anything
Year: 2024
Type: article
Abstract: Recent advancements in deep learning have enabled its application to real-world scenarios, yet many deep learning models face challenges in real-time processing due to high computational demands. This study proposes a method for real-time image transformation that emphasizes the boundary between foreground and background while addressing these limitations. The proposed approach utilizes the Depth Anything model to generate a depth-based background image and identify the Region of Interest (RoI) from the input image using GPU. Subsequently, the RGB input image is transformed into HSV color space, and a final image is created by combining the value channel, an inverted saturation channel, and the depth background image. Once the depth background is initialized, subsequent image transformations can be performed in real-time on a CPU. Experimental results demonstrate that the proposed method highlights foreground-background boundaries comparably to Depth Anything while significantly reducing computational overhead. The method is particularly effective for video monitoring applications with static cameras, ensuring computational efficiency and robustness.
Authors: Jongwoong Seo, Chewon Kim, Seungwoo Lee, Yongwha Chung
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4406747534