Generating Superpixels for High-Resolution Images with Decoupled Patch Calibration
Superpixel segmentation is a significant task in the field of computer vision that involves grouping pixels with similar attributes into coherent clusters known as superpixels.These superpixels are not only useful for image annotation but also serve as a foundation for various downstream applications such as segmentation,optical flow estimation,and depth estimation.Despite the substantial progress in superpixel segmentation techniques,particularly with the advent of deep learning methods,existing solutions have been unable to effectively handle high-resolution images due to constraints in GPU memory and computational power.The authors propose the Patch Calibration Network(PCNet),a novel deep learning framework that addresses the limitations of current methods by employing a decoupled consistency learning strategy.This approach allows for the efficient generation of high-resolution superpixels by predicting high-resolution outputs from low-resolution inputs,thereby bypassing the GPU memory limitations.A key aspect of PCNet is the Decoupled Patch Calibration(DPC)branch,which incorporates high-resolution image patches as additional inputs to preserve fine details and enhance boundary pixel allocation.To improve the identification of boundary pixels,the authors introduce a dynamic guidance training mechanism that utilizes a binary mask.This mechanism encourages the network to focus on the primary boundaries within a region,simplifying the task from multi-class classification to a binary classification problem.This innovative strategy not only reduces the complexity of network optimization but also significantly enhances the precision of boundary detection.The paper demon-strates the effectiveness of PCNet through extensive experiments on diverse datasets,including Mapillary Vistas,BIG,and a newly created Face-Human dataset.The results indicate that PCNet can successfully process 5K resolution images and achieve superior performance compared to the state-of-the-art SCN method,which struggles with high-resolution inputs.The authors'contributions include the development of PCNet,a deep learning solution for high-resolution superpixel segmentation,the introduction of a decoupled regional calibration architecture,and the construction of an ultra-high-resolution benchmark dataset for evaluating the performance of superpixel segmentation algorithms in high-resolution scenarios.The paper is structured to first review the related work in the field of superpixel segmentation,then present the PCNet framework in detail,followed by experimental results and comparisons with state-of-the-art methods.The conclusion summarizes the findings and outlines potential directions for future research.The availability of code,pre-trained models,and the new benchmark dataset will undoubtedly facilitate further advancements in the field of high-resolution superpixel segmentation.In summary,this paper presents a significant advancement in the domain of superpixel segmentation,providing a solution that can handle high-resolution images efficiently and accurately.The proposed PCNet framework,with its innovative DPC branch and dynamic guidance training mechanism,offers a promising direction for future research and applications in computer vision.Our code,pre-trained models,and the newly constructed evaluation benchmark dataset are available at https://github.com/wangyxxjtu/PCNet.
万方数据
维普
