立體視覺之低階運算與高階調控

The Interaction of Low-Level Computation and High-Level Modulation in Human Stereopsis

本文評述在1983-2000年間於本實驗室所進行之立體視覺研究，以測試Marr（1982）所帶動之計算視覺理論。本文先綜述計算視覺理論中之兩大基石，亦即「空間對應假設」及「立體視覺基本假設」，以及可供有效電腦模擬的「合作算則」；之後依「計算視覺理論及其測試」、「神經機制與雙眼像差之獨立性與優先性」、及「立體視覺中的高階互動效應」三項子題，驗證Marr（1982）與傳統計算理論中有關視覺回復與反向光學議題。合作算則中的「相容性」限制條件，並未能切合雙眼配對之實驗資料；利用兩類眼動模組為基礎，以方位訊息為配對元素之模擬，應可成功取代合作算則對不同像差之計算。雙眼像差及錯覺輪廓因有明確的神經機制，故具有影響知覺結果之獨立性與優先性，其處理先後似可依序排列為像差→輪廓→顏色及其他次級知覺，惟尚有待進一步驗證。但該事實並不能排除高階調控之影響，在形狀與不確定像差對立體攫取之影響、漢字背景對錯覺輪廓之約束力、單眼線索對像差運作之影響，以及盲點區之深度填補等項上，皆可看到高階因素介入互動之可能。本文最後主張視覺計算理論應將回饋機制列入考量，以有效處理立體視覺中低階運算與高階調控之互動。

The present article reviews a set of studies that were conducted in the period of 1983-2000 for a test of the evolving computational vision theory after the posthumous publication of Marr's (1982) classic on vision. Two pillars of the main architecture, namely, the spatial correspondence assumption and the fundamental assumption of stereopsis, with their realization into a successful implementation of cooperative algorithm were briefly described and put into a series of critical tests. A summary evaluation was suggested to highlight the plausibility and inadequacies concerning an application of retinal matching constraints (especially, the constraint of compatibility) in human stereoscopic processing. An alternative approach based on the operation of conjunctive and vergent eye movements in finding hidden disparities has been simulated to show its successful resolution of visual recovery, and henceforth, inverse optics. Further experimental evidence were provided to show the operational precedence of disparity and contour information with a compromise to the modulation of high-level factors in human stereoscopic processing. High-level influence could be brought into effect to resist the stereo capture to a positive disparity plane where residing the illusory contour while a Chinese character was resting on a zero-disparity background. A similar result was also observed in a stereoscopic experiment on bistable figure. Monocular cues were shown to be effectively modulating the disparity computation. The study also maintains that a cognitive interpretation can not be excluded in the discussion of disparity interpolation in the blind-spot area. The study therefore proposes that an evolving computational vision theory should better take feedback mechanism into account to deal with the intricate interaction of low-level computation and high-level modulation.