Visual target detection under multiple dimensions of uncertainty

Detection of visual targets is integral to survival and everyday functioning. In the real world, the visual system operates under very high levels of extrinsic uncertainty (multiple simultaneous dimensions of uncertainty) about target and background properties. However, the previous literature is primarily concerned with the effect of low and modest levels of uncertainty (only a single dimension of uncertainty) on human performance. This thesis aims to measure and model human performance under multiple simultaneous dimensions of uncertainty. I primarily focus on the detection of additive targets in white noise. First, human performance was measured under simultaneous target amplitude, and background contrast uncertainty for target prior probabilities of 0.5 and 0.2. I simulated and tested three model observers. The ideal observer, which has a dynamic decision criterion, can be approximated with a contrast normalized template-matching (NTM) observer with a single criterion. Maximum-likelihood fits revealed that the NTM and the dynamic-decision-criterion observer predict human performance much better than a template-matching (TM) observer with a single criterion. The same results were also found for natural backgrounds. The results reveal the value contrast normalization under real-world conditions where target priors are low. Secondly, human performance was measured under simultaneous target scale and target orientation uncertainty. I also describe an efficient simulation method. Simulations revealed that the maximum-template-response (MAX) observer only approximates the ideal observer when template responses are normalized by the energy of templates (ENM observer). Maximum likelihood fits reveal the ENM and the ideal observer explain human performance much better than the simple MAX observer. Furthermore, human performance was also measured under low uncertainty. I found that humans are more efficient under high uncertainty. The efficiency difference can be accounted for by incorporating intrinsic position uncertainty into the model observers. The results reveal the value of energy normalization and the importance of intrinsic uncertainties for understanding the visual detection under various levels of uncertainty. This thesis provided insights into visual processing under uncertainty. It exemplifies the fruitfulness of studying detection under simultaneous multiple dimensions of uncertainty, within a principled framework. The methods developed in this this should be useful in future experiments.