GHOF-Cam
收藏资源简介:
GHOF-Cam是一个专门为相机运动估计设计的基准数据集,由现有的光流数据集中动态物体和不良遮挡区域的掩码创建而成,旨在隔离纯相机运动。该数据集通过使用Segment Anything Model (SAM)生成语义图,手动识别动态对象(如汽车和人)并创建相应的掩码,然后将掩码应用于输入图像和真实光流,从而仅捕获相机引起的运动。
GHOF-Cam is a benchmark dataset specifically designed for camera motion estimation. It is constructed using masks of dynamic objects and poorly occluded regions extracted from existing optical flow datasets, with the goal of isolating pure camera-induced motion. This dataset is developed by first generating semantic maps via the Segment Anything Model (SAM), manually identifying dynamic objects such as cars and humans and creating corresponding masks, then applying these masks to both input images and ground-truth optical flow, thereby only capturing motion caused by camera movement.
CamFlow数据集概述
基本信息
- 数据集名称: Estimating 2D Camera Motion with Hybrid Motion Basis (CamFlow)
- 作者: Haipeng Li, Tianhao Zhou, Zhanglei Yang, Yi Wu, Yan Chen, Zijing Mao, Shen Cheng, Bing Zeng, Shuaicheng Liu
- 机构: 电子科技大学, 小米公司, Dexmal
- 项目页面: https://lhaippp.github.io/CamFlow/
研究背景
- 核心问题: 传统基于单应性的方法无法表示组合流场产生的复杂非线性变换
- 关键发现: 组合不同单应性产生的流场会形成无法用任何单一单应性表示的运动模式
方法创新
- 混合运动基:
- 物理运动基(基于相机几何原理)
- 随机运动基(处理复杂非线性变换)
- 技术组件:
- 运动估计变换器(MET)
- 不确定性预测模块
- 基于拉普拉斯分布的混合概率损失函数
数据集特点
- 构建方式: 通过掩蔽现有光流数据集中的动态物体来隔离纯相机运动
- 场景覆盖:
- 低光照环境
- 雨雪天气
- 雾天环境
实验结果
定量评估
CAHomo Benchmark (PME)
- 最佳方法: Ours (0.32)
- 对比方法: SIFT + MAGSAC (1.34), DMHomo (0.31)
GHOF-Cam (EPE)
- 最佳方法: Ours (1.10)
- 对比方法: BasesHomo (1.74), Meshflow (2.15)
GHOF Test (PME)
- 最佳方法: Ours (1.23)
- 对比方法: RealSH (1.72), DMHomo (1.75)
综合评估(PSNR/SSIM/LPIPS)
| 指标 | Ours | 对比最佳方法 |
|---|---|---|
| PSNR↑ | 32.09 | GT-Homo(32.78) |
| SSIM↑ | 0.9142 | GT-Homo(0.9187) |
| LPIPS↓ | 0.0575 | GT-Homo(0.0570) |
引用格式
bibtex @inproceedings{li2025estimating, title={Estimating 2D Camera Motion with Hybrid Motion Basis}, author={Li, Haipeng and Zhou, Tianhao and Yang, Zhanglei and Wu, Yi and Chen, Yan and Mao, Zijing and Cheng, Shen and Zeng, Bing and Liu, Shuaicheng}, booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision}, pages={xxxx--xxxx}, year={2025} }




