在图像生成中引入可调整大小的区域打包问题及其启发式求解方法
Introducing Resizable Region Packing Problem in Image Generation, with a Heuristic Solution
AI总结：
本文在合成图像生成中提出可调整大小的锚定区域打包（RARP）问题并推测其NP难，进而设计了一种通用贪心启发式算法，可高效打包任意数量与形状的区域，并通过大规模异常检测数据集生成实验验证了其有效性。
摘要：
计算机视觉中的图像数据生成问题历来比判别性问题更难解决。此类数据生成需要在场景画布中的有意义位置放置尺寸适当的多个相关对象。目前流行的生成方法主要有两类：基于图形的方法和基于生成模型的方法。已知在这两类方法的背后都潜藏着优化问题。本文在合成图像数据生成的背景下，引入了经典装箱问题的一种新颖且具有实际用途的表现形式。我们推测这一新引入的问题——可调整大小的锚定区域打包（Resizable Anchored Region Packing，RARP）问题——是NP难的，并给出了关于该推测的详细论证。作为首个解决方案，我们提出了一种新颖的启发式算法，该算法具有足够的通用性，因此能够将任意数量、任意形状的区域在任意位置进行缩放并打包到图像画布中。该算法采用贪心策略，在遵守优化约束的同时，以精细的方式迭代地打包区域对。通过一个实现验证了该算法，该实现用于生成大规模合成异常检测数据集，每个图像样本（即RARP实例）的装箱参数变化程度很高。对此类数据的目视检查以及对每个解的正确性检验证明了我们算法的有效性。随着生成建模在深度学习中的兴起，以及合成数据生成有望成为主流，我们预期这一新引入的问题将受到成像科学界的重视。
英文摘要：
The problem of image data generation in computer vision has traditionally been a harder problem to solve, than discriminative problems. Such data generation entails placing relevant objects of appropriate sizes each, at meaningful location in a scene canvas. There have been two classes of popular approaches to such generation: graphics based, and generative models-based. Optimization problems are known to lurk in the background for both these classes of approaches. In this paper, we introduce a novel, practically useful manifestation of the classical Bin Packing problem in the context of generation of synthetic image data. We conjecture that the newly introduced problem, Resizable Anchored Region Packing(RARP) Problem, is NP-hard, and provide detailed arguments about our conjecture. As a first solution, we present a novel heuristic algorithm that is generic enough and therefore scales and packs arbitrary number of arbitrary-shaped regions at arbitrary locations, into an image canvas. The algorithm follows greedy approach to iteratively pack region pairs in a careful way, while obeying the optimization constraints. The algorithm is validated by an implementation that was used to generate a large-scale synthetic anomaly detection dataset, with highly varying degree of bin packing parameters per image sample i.e. RARP instance. Visual inspection of such data and checking of the correctness of each solution proves the effectiveness of our algorithm. With generative modeling being on rise in deep learning, and synthetic data generation poised to become mainstream, we expect that the newly introduced problem will be valued in the imaging scientific community.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
