Author:

Keywords:

context-based reasoning, object detection, object pose estimation, collective classification, relational learning, PSI_VISICS, PSI_3900

Abstract:

Computer vision algorithms have become very eff ective at detecting the occurrence of objects in i mages. Parallel to this, notable advancements have been achieved in estimating the orientation¨ in which such objects occur. Different methods hav e been proposed during the last decade, ranging fr om methods that model the appearance of the object ¨as it is projected on the 2D image space to metho ds that reason about physical properties of the ob jects in the 3D scene. These methods have pro ved to be effective at the tasks at hand. How ever, one weakness of these methods is their compl ete reliance on intrinsic features, e.g. colo r, size, texture, that define the objects of inter est. This weakness becomes evident in difficult sc enarios triggered by factors such as high int er-object occlusion; which affects perceived shape ¨and size of objects, as well as drastic chan ges in illumination; which affects how the texture ¨and color of objects are perceived by the camera. There are additional, extrinsic, cues that¨ can help under these scenarios. For example, some¨ object categories tend to appear more often in som e scene types than in others. For instance, i t is more likely to find a computer in an ind oor scene rather than in an outdoor setting. Likew ise, in natural and man-made objects there are som e, imposed or desired, rules that determine the co nfigurations in which objects co-occur. For e xample, birds fly following a flocking behavior, k eyboard and mouse are usually found below the computer screen, and so on. This thesis investiga tes the potential of these extrinsic cues to assis t computer vision tasks such as object detection a nd object pose estimation. Context cues have been used before for object detection. Here¨ we show that they can also help in object pose est imation. This applies to both scene cues (e.g. the ¨groundplane) as well as location and pose of othe r objects in the scene. Furthermore, we show that¨ cautious inference on object relations brings impr ovements over traditional inference for object det ection. Finally, we show how to use context cues not only to filter our false object detections but¨ also to retrieve object instances missed in a n initial detection step.