1 Schedule of Talks

In the task of visual object categorization, semantic context can play the very important role of reducing ambiguity in objects' visual appearance. However, the majority of newly proposed object recognition models -whether generative or discriminative – only consider the intrinsic object features. In this work we introduce a novel approach to object categorization that incorporates two types of context -co-occurrence and relative location -with local appearance-based features. Our approach, named CoLA (for Co-occurrence, Location and Appearance), uses a conditional random field (CRF) to maximize object label agreement according to both semantic and spatial relevance. We compare two sources of semantic context: one learned from training data and another queried from Google Sets. To capture spatial constrains, we model relative location between objects using simple pairwise features. By vector quantizing this feature space, we learn a small set of prototypical spatial relationships directly from the data. The overall performance of the proposed framework is evaluated on the PASCAL and MSRC datasets. Our findings conclude that incorporating context into object categorization greatly improves categorization accuracy. We evaluate our results on two challenging datasets: PASCAL 2007 and MSRC. The results also show that combining co-occurrence and spatial context improves accuracy in as many as half of the categories compared to using co-occurrence alone. A multinomial processing tree model of change blindness and change detection Emmanuelle Boloix Massachusetts Institute of Technology In most change blindness studies, the failure to report changes is attributed to either a representational failure (the pre-change scene is not represented in memory) or to a comparison failure (preand postchange scenes are not compared). Here, I propose a Multinomial Processing-Tree (MPT) model, which determines the relative contributions of representational and comparison failures to change blindness. MPT models are statistical models used to measure latent cognitive processes from observable raw data. Cognitive processes are represented as model parameters, their respective weights can be assessed, and the fit of the model to the empirical data can be evaluated via goodness-of-fit tests (Batchelder & Riefer, 1999). The MPT model I propose assumes that visual information can be either represented in memory or non-represented, and, if represented, that visual representation can be compared to the currently-displayed view or not. Model parameters have been set to be consistent with various theories of scene representation (Hollinworth, 2004; Irwin & Zelinski, 2002; Rensink, 2000; Triesch et al, 2003). Here, I ran one change blindness experiment to test the model’s predictions. Observers (N=132) were shown a map of a virtual village and had to perform a 7-stage route in the map. Changes occurred on task-relevant or task-irrelevant objects. Observers were instructed to complete the route as quickly as possible, and to report perceived changes. Our results showed that observer’s change detection performance was best modelled when MPT parameters were set according to a Task-Relevance theory of scene representation. We then used the model to quantify how much of change blindness is due to a comparison failure or to a representational failure. Results show that differences in change detection performances are due to the quality of representation in memory, which depends on object’s relevance for the task, and not to a comparison process (the model’s weighting of the memory representation parameter varies from 19% to 82%, whereas the weighting of the memory comparison parameters remains stable in the range from 28% to 38%). Implications for several theories of scene representation are discussed.