Search

Activity interpretation in videos results not only in recognition or labeling of dominant activities, but also in semantic descriptions of scenes. Towards this broader goal, we present a combinatorial approach that assumes availability of algorithms for detecting and labeling objects and basic actions in videos, albeit with some errors. Given these uncertain labels and detected objects, we link them into interpretable structures using the domain knowledge, under the framework of Grenander's general pattern theory. Here a semantic description is built using basic units, termed generators, that represent either objects or actions. These generators have multiple out-bonds, each associated with different types of domain semantics, spatial constraints, and image evidence. The generators combine, according to a set of pre-defined combination rules that capture domain semantics, to form larger configurations that represent video interpretations. This framework derives its representational power from flexibility in size and structure of configurations. We impose a probability distribution on the configuration space, with inferences generated using a Markov chain Monte Carlo-based simulated annealing process. The primary advantage of the approach is that it handles known challenges-appearance variabilities, errors in object labels, object clutter, simultaneous events, etc-without the need for exponentially-large (labeled) training data. Experimental results demonstrate its ability to successfully provide interpretations under clutter and the simultaneity of events. They show: (1) a performance increase of more than 30 % over other state-of-the-art approaches using more than 5000 video units from the Breakfast Actions dataset, and (2) an overall recall and precision improvement of more than 50 and 100 %, respectively, on the YouCook data set. [ABSTRACT FROM AUTHOR]

• We propose a Transformer-based network architecture for video salient object detection. • We propose a Gated Cross Reference (GCR) module to facilitate collaborative learning between appearance and motion cues. • The proposed Transformer-based method reaches excellent performance on five public benchmarks. Video salient object detection is a fundamental computer vision task aimed at highlighting the most conspicuous objects in a video sequence. There are two key challenges presented in video salient object detection: (1) how to extract effective feature representations from appearance and motion cues, and (2) how to combine both of them into robust saliency representation. To handle these challenges, in this paper, we propose a novel Transformer-based Cross Reference Network (TCRN), which fully exploits long-range context dependencies in both feature representation extraction and cross-modal (i.e., appearance and motion) integration. In contrast to existing CNN-based methods, our approach formulates video salient object detection as a sequence-to-sequence prediction task. In the proposed approach, the deep feature extraction is achieved by a pure vision transformer with multi-resolution token representations. Specifically, we design a Gated Cross Reference (GCR) module to effectively integrate appearance and motion into saliency representation. The GCR first propagates global context information between different modalities, and then perform cross-modal fusion by a gate mechanism. Extensive evaluations on five widely-used benchmarks show that the proposed Transformer-based method performs favorably against the existing state-of-the-art methods [ABSTRACT FROM AUTHOR]