Tianshi Gao

Discriminative learning of relaxed hierarchy for large-scale visual recognition

In the real visual world, the number of categories a classifier needs to discriminate is on the order of hundreds or thousands. For example, the SUN dataset [24] contains 899 scene categories and ImageNet [6] has 15,589 synsets. Designing a multiclass classifier that is both accurate and fast at test time is an extremely important problem in both machine learning and computer vision communities. To achieve a good trade-off between accuracy and speed, we adopt the relaxed hierarchy structure from [15], where a set of binary classifiers are organized in a tree or DAG (directed acyclic graph) structure. At each node, classes are colored into positive and negative groups which are separated by a binary classifier while a subset of confusing classes is ignored. We color the classes and learn the induced binary classifier simultaneously using a unified and principled max-margin optimization. We provide an analysis on generalization error to justify our design. Our method has been tested on both Caltech-256 (object recognition) [9] and the SUN dataset (scene classification) [24], and shows significant improvement over existing methods.

A segmentation-aware object detection model with occlusion handling

The bounding box representation employed by many popular object detection models [3, 6] implicitly assumes all pixels inside the box belong to the object. This assumption makes this representation less robust to the object with occlusion [16]. In this paper, we augment the bounding box with a set of binary variables each of which corresponds to a cell indicating whether the pixels in the cell belong to the object. This segmentation-aware representation explicitly models and accounts for the supporting pixels for the object within the bounding box thus more robust to occlusion. We learn the model in a structured output framework, and develop a method that efficiently performs both inference and learning using this rich representation. The method is able to use segmentation reasoning to achieve improved detection results with richer output (cell level segmentation) on the Street Scenes and Pascal VOC 2007 datasets. Finally, we present a globally coherent object model using our rich representation to account for object-object occlusion resulting in a more coherent image understanding.

User-centric speaker report: Ranking-based effectiveness evaluation and feedback

In this paper, we present a computer vision-based system that is capable of automatically analyzing presentation videos and evaluating it according to the learned users preference. In this system, different visual features indicating the effectiveness of the presentation are extracted. They include the speakers global movement, face/head orientation distribution and motions caused by the use of hands. Given a set of user scored presentation videos, we adapt the RankBoost [13] algorithm to learn the users scoring preference so that the system can score a new presentation video in the future to provide the user feedback. The experiment results show that the vision processing part can reliably extract the low level features and the ranking learning part can successfully learn users different scoring preferences and achieve an average ranking error within one level or less.

Self lane assignment using egocentric smart mobile camera for intelligent GPS navigation

In this paper, we study the self lane assignment problem, i.e. given an image taken inside a vehicle, infer on which lane the image is taken. This problem serves as an example of active egocentric vision application with data fusion. In this application, a camera is mounted inside the vehicle looking outside to the world. Combined with a GPS with a digital map this smart mobile camera is capable of reasoning on which lane the vehicle is. This inference result is then fed back to the GPS to provide the driver with more intelligent navigation instructions. We form the self lane assignment inference problem as a scene classification problem which requires classifying scenes in finer categories than the traditional case. We design the features to represent the image in a holistic way bypassing individual object detection, develop an automatic horizon detection algorithm, and employ and compare three learning algorithms for decision making on the lane number. The experiment results show that our method can achieve the precision and recall rates around or above 90% at the same time.

Load-Balanced Cluster-Based Cooperative MIMO Transmission for Wireless Sensor Networks

Remote environmental surveillance by massively deployed tiny wireless sensors requires energy efficient communication and network protocols so as to prolong the network lifetime. A load-balanced cluster-based cooperative MIMO (multiple-in-multiple-out) transmission scheme for such wireless sensor networks is proposed, taking imperfect data aggregation into consideration. In this scheme, a two-layer hierarchy is formed by clustering, and the cluster heads perform local data aggregation, balance communication loads and transmit data back to the base station using cooperative MIMO techniques. Simulation results show that the proposed scheme can distribute the energy dissipation more evenly throughout the network and achieve higher energy efficiency, which leads to a longer network life span compared with other traditional schemes.