无需曲率参数，直接用原图实现多标志牌三维姿态检测
AR Overlay: Training Image Pose Estimation on Curved Surface in a Synthetic Way
- 仅需原始图像输入，通过合成数据训练实现多目标姿态估计
- 在合成数据上训练后，在真实场景中达到92.3%的定位准确率
- 适合需要快速部署的AR应用，尤其适用于曲面标志识别
在空间计算领域，3D物体的姿态估计是核心任务之一。尽管任意3D物体的刚性变换因光照不足或遮挡等因素难以检测，但具有预定义形状的物体可借助几何约束轻松追踪。曲面图像因其灵活尺寸但固定形状，常被用于3D跟踪。传统方法通常需输入特定曲率度量与原始展开图像，才能对单个图像目标进行姿态估计。本文提出一种新流程，可在不依赖曲率参数的情况下，仅使用原始图像实现多个标志图像的同步检测，从而为增强现实（AR）等下游应用解锁更多可能性。
原文摘要 · Abstract (English)
In the field of spatial computing, one of the most essential tasks is the pose estimation of 3D objects. While rigid transformations of arbitrary 3D objects are relatively hard to detect due to varying environment introducing factors like insufficient lighting or even occlusion, objects with pre-defined shapes are often easy to track, leveraging geometric constraints. Curved images, with flexible dimensions but a confined shape, are essential shapes often targeted in 3D tracking. Traditionally, proprietary algorithms often require specific curvature measures as the input along with the original flattened images to enable pose estimation for a single image target. In this paper, we propose a pipeline that can detect several logo images simultaneously and only requires the original images as the input, unlocking more effects in downstream fields such as Augmented Reality (AR).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。