Citation Context

...now briefly describe previous work in each of these areas. Perceiving the environment. Perception is a key element of many robotic tasks. It has been applied to object labeling [37, 2, 60], scene segmentation [23], robot localization [43, 46], feature extraction for planning and manipulation [30], understanding environment constraints [28] and object affordances [12, 34]. Sharing representations from visual information not only improve the performance of each of the perception tasks, but also significantly help various applications such as autonomous or assistive driving [14, 8], anticipation [31, 35, 59], planning sociable paths [36] and for various household chores such as grasping and cutting [41]. Sharing representations from other modalities such as sound [52] and haptics [21] would also improve perception. Path planning and manipulation. There exist a large class of algorithms which allow robots to move around and modify the environment. Broadly the planning algorithms can be categorized as motion planning [62, 53], task planning [1, 7] and symbolic planning [17, 51]. Bakebot [7], towel-folding [54], IkeaBot [32] and robots preparing pancakes [5] are few of the many successful planning a...

Citation Context

...olute motion of the boy, and (e) the relative motion of the boy and the cap. Notice it is more likely for the hat to move with the boy than not. modeled. This difficulty has also noted in recent work =-=[17]-=- that, in addition to prediction, tackles the challenge of learning which parts of the scene will move without assuming apriori semantic knowledge. Nonetheless, as confirmed by our experiments, the mo...

Citation Context

...a sequence of images [31, 23, 15, 8, 17]. In contrast, we wish to deduce the motivation of actions in a single image, and not what will happen next. There also has been work in forecasting activities =-=[16, 32]-=-, inferring goals [35], and early event detection [11], but they are interested in predicting the future in videos while we wish to explain the motivations of actions of people in images in the wild. ...

Citation Context

...bits, while EDD would learn the identity map, since the data distributions are the same at any time. A related line of existing work aims at predicting the future motion of specific objects in videos =-=[11, 25, 28, 29]-=-, or anticipating the behavior of a human in order to facilitate interaction with a robot [12, 26]. Such model-based approaches rely on task-specific assumptions, e.g. on the visual appearance of obje...

Citation Context

...ause they are cheap to obtain at massive scales and still contain rich signals. Foundational work in computer vision has explored the challenging task of visualizing the future using unlabeled videos =-=[29, 45, 48]-=-. However, in many applications, we are interested in predicting the presence of concepts in the future, which may be easier than predicting low level signals. Rather than predicting pixels, the main ...

Citation Context

...n a 2D image. In order tosaddress the challenge posed by question 2, we look to the results generated by Walker, Guptasand Herbert in their paper “Patch to the Future: Unsupervised Visual Prediction” =-=[72]-=-. In thisspaper, the authors utilized a reward function on textures that are to be traversed by the IOI withinsthe video frame and subsequently generated a set of high probability routes to be followe...

Citation Context

...mages or video, here we address a different task and predict actions in scene images with no action observations. Action prediction has been recently addressed by Kitani et al. [10] and Walker et al. =-=[16]-=- aiming to model future motion of people and cars using priors derived from the scene. Yuen and Torralba [17] predict motion for images of static scenes by searching and transferring motion cues from ...