“School of Cognitive Sciences”
Back to Papers HomeBack to Papers of School of Cognitive Sciences
| Paper IPM / Cognitive Sciences / 18356 |
|
||||||
| Abstract: | |||||||
|
Scene context is known to significantly influence visual perception, enhancing object recognition particularly under challenging viewing conditions. Behavioral and neuroimaging studies suggest that high-level scene information modulates activity in object-selective brain areas through top-down mechanisms, yet the underlying mechanism of this process remains unclear. Here, we introduce a biologically inspired context-based computational model (CBM) that integrates scene context into object recognition via an explicit feedback mechanism. CBM consists of two distinct pathways: Object_CNN, which processes localized object features, and Place_CNN, which extracts global scene information to modulate object processing. We compare CBM to a standard feedforward model, AlexNet, in a multiclass object recognition task under varying levels of visual degradation and occlusion. CBM significantly outperformed a standard feedforward model (AlexNet), demonstrating the effectiveness of structured contextual feedback in resolving ambiguous or degraded visual input. However, behavioral experiments revealed that while humans also benefited from congruent context - particularly at high occlusion levels - the effect was modest. Human recognition remained relatively robust even without contextual support, suggesting that mechanisms such as global shape processing and pattern completion, likely mediated by local recurrent processes, play a dominant role in resolving occluded input. These findings highlight the potential of contextual feedback for enhancing model performance, while also underscoring key differences between human and models. Our results point toward the need for models that combine context-sensitive feedback with object-intrinsic local recurrent processes to more closely approximate the flexible and resilient strategies of human perception
Download TeX format |
|||||||
| back to top | |||||||
