FUSION AND CALIBRATION OF MULTIPLE DEPTH CUES

生物视觉系统利用几种深度线索（遮挡， 纹理、透视、运动、视差）。 我们提出了联合收割机 计算这些线索的模块的结果。 我们描述了一个 理想的深度观察者，它结合了单独的深度线索，以绝对 深度和加权线索相对于其估计的可靠性。 的 分配给线索的标准权重应该随着场景和观看而变化 条件（例如，场景中的纹理数量）。 一个辅助线索 用于评估各种深度模块的可能性能。 我们 检查辅助线索和权重之间的（复杂）映射 选择的规则的组合分析，通过比较与 人类心理物理性能，并通过自适应网络 模拟 我们研究了两种类型的学习：校准和深度融合 学习 校准将各种模块的输出转换为 真实的深度估计。 深度融合学习开发了一个映射， 辅助线索值到最佳线索权重。 我们将开发：（1） 心理物理测量的深度组合规则使用的人 观察员当线索（近似）在和谐;（2）一个软件 理想和心理物理深度的模拟和建模试验台 观察者;（3）基于这些数据的心理物理观察者模型 和规范（“理想的”）模型;（4）校准模型 深度融合学习 这项拟议的研究将使我们能够进一步了解 人类视觉系统的多重深度提示。 的理解 校准过程可立即应用于重新校准， 发生在生物视觉中，当基本参数随时间变化时， 例如瞳孔间距。 融合学习的研究将 揭示了人类视觉系统如何使 随着视觉器官的变化（通过 老化和/或疾病），从而改变 线索

Biological visual systems make use of several depth cues (occlusion, texture, perspective, motion, disparity). We propose methods to combine the results of the modules which compute these cues. We describe an ideal depth observer which combines the separate depth cues to absolute depth, and weighting cues relative their estimated reliability. The normative weights assigned to cues should vary with the scene and viewing conditions (e.g., he amount of texture in the scene). An ancillary cue is used to assess the likely performance of various depth modules. We examine the (complicated) mapping between ancillary cues and weights selected for the rule of combination analytically, by comparison with human psychophysical performance, and through adaptive network simulations. We investigate two types of learning; calibration, and depth fusion learning. Calibration translates the output of various modules to veridical depth estimates. Depth fusion learning develops a mapping from ancillary cue values to optimal cue weights. We will develop: (1) psychophysical measurements of the depth combination rule used by human observers when cues are (approximately) in harmony; (2) a software testbed for the simulation and modeling of ideal and psychophysical depth observers; (3) models of the psychophysical observer based on these data and normative ('ideal observer') models; and (4) models of calibration and depth fusion learning. The proposed research will allow us to further understand the use of multiple depth cues by the human visual system. An understanding of the calibration process is immediately applicable to the recalibration that takes place in biological vision when basic parameters change over time, such as interpupillary distance. The research on fusion learning will shed light on how the human visual system can make the most reliable estimates of depth possible as the visual apparatus changes (through aging and/or disease) so as to alter the relative reliability of the cues.