Mid-level vision and recognition of non-rigid objects

In this dissertation I address the problem of visual recognition of non rigid objects I introduce the frame alignment approach to recognition and illustrate it in two types of non rigid objects contour textures and elongated exible objects Frame alignment is based on matching stored models to images and has three stages rst a frame curve and a corresponding object are computed in the image Second the object is brought into correspondence with the model by aligning the model axis with the object axis if the object is not rigid it is unbent achieving a canonical description for recognition Finally object and model are matched against each other Rigid and elongated exible objects are matched using all contour information Contour textures are matched using lter outputs around the frame curve The central contribution of this thesis is Curved Inertia Frames C I F a scheme for computing frame curves directly on the image C I F is the rst algorithm which can compute probably global and curved lines and is based on three novel concepts rst a de nition of curved axis of inertia second the use of non cartisean networks third a ridge detector that automatically locates the right scale of objects The use of the ridge detector enables C I F to perform mid level tasks without the need of early vision C I F can also be used in other tasks such as early vision perceptual organization computation of focus of attention and part decomposition I present evidence against frame alignment in human perception However this ev idence suggests that frame curves have a role in gure ground segregation and in fuzzy boundaries and that their outside near top incoming regions are more salient These ndings agree with a model in which human perception begins by setting a frame of refer ence prior to early vision and proceeds by successive processing of convex structures or holes The schemes presented for contour texture and elongated exible objects use a comon two level representation of shape and contour texture which may also be useful to recog nize other non rigid transformations such as holes rigid objects articulated objects and symbolic objects Most of the schemes have been tested on the Connection Machine and compared against human perception Some work on part segmentation and holes is also presented Thesis Committee Prof W Eric L Grimson Prof Tomas Lozano Perez Prof Tomaso Poggio Prof Whitman Richards Prof Shimon Ullman