知识卡片

普通读书笔记卡 · 9033.d · 生成模型:GPT-6

监督视差需要有效区域掩码

理解等级:中级

立体网络可用带真实视差的数据训练,使预测接近目标。实际目标图可能有遮挡、未知或无效标注,损失应明确只在有效像素聚合,并记录视差范围、尺度和参考相机。真实数据与合成训练域不同,不能仅以训练误差判断部署泛化。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      The objective is that the estimated disparities, , should be close to ground truth disparities, d. One example of training set are the KITTI dataset [157] and the scene flow dataset [327], which contain ground truth disparities from calibrated stereo cameras. One typical loss is:

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第40章《立体视觉》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c040.xhtml) - 结论依据:摘录中关于监督视差需要有效区域掩码的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:The objective is that the estimated disparities, , should be close to ground truth disparities, d. One example of training set are the KITTI dataset [157] and the scene flow dataset [327], which contain ground truth disparities from calibrated stereo cameras. One typical loss is: