Search

Object recognition and localization are important to automatically interpret video and allow better querying on its content. We propose a method for object localization that learns incrementally and addresses four key aspects. Firstly, we show that for certain applications, recognition is feasible with only a few training samples. Secondly, we show that novel objects can be added incrementally without retraining existing objects, which is important for fast interaction. Thirdly, we show that an unbalanced number of positive training samples leads to biased classi er scores that can be corrected by modifying weights. Fourthly, we show that the detector performance can deteriorate due to hard-negative mining for similar or closely related classes (e.g., for Barbie and dress, because the doll is wearing a dress). This can be solved by our hierarchical classi cation. We introduce a new dataset, which we call TOSO, and use it to demonstrate the e ectiveness of the proposed method for the localization and recognition of multiple objects in images., This research was performed in the GOOSE project, which is jointly funded by the enabling technology program Adaptive Multi Sensor Networks (AMSN) and the MIST research program of the Dutch Ministry of Defense. This publication was supported by the research program Making Sense of Big Data (MSoBD)., peer-reviewed

We introduce a canonical representative for the isomorphism classes of non-regular orthogonal arrays based on the generalized word length patterns (GWLP) of delete-one-factor projections. These GWLP values have been used recently to introduce a fast isomorphism test for two-level regular arrays. We show that the delete-one-factor projection GWLP method can be adopted to reduce both regular and non-regular orthogonal arrays to canonical form. The new canonical form is used in an existing framework to generate minimal complete sets of non-symmetric non-regular arrays. We show that the new method is efficient for reduction to canonical form, but not suitable for generating minimal complete sets.

We describe the TNO system and the evaluation results for TRECVID 2013 Multimedia Event Detection (MED) and instance search (INS) tasks. The MED system consists of a bag-of-word (BOW) approach with spatial tiling that uses low-level static and dynamic visual features, an audio feature and high-level concepts. Automatic speech recognition (ASR) and optical character recognition (OCR) are not used in the system. In the MED case with 100 example training videos, support-vector machines (SVM) are trained and fused to detect an event in the test set. In the case with 0 example videos, positive and negative concepts are extracted as keywords from the textual event description and events are detected with the high-level concepts. The MED results show that the SIFT keypoint descriptor is the one which contributes best to the results, fusion of multiple low-level features helps to improve the performance, and the textual event-description chain currently performs poorly. The TNO INS system presents a baseline open-source approach using standard SIFT keypoint detection and exhaustive matching. In order to speed up search times for queries a basic map-reduce scheme is presented to be used on a multi-node cluster. Our INS results show above-median results with acceptable search times., This research for the MED submission was performed in the GOOSE project, which is jointly funded by the enabling technology program Adaptive Multi Sensor Networks (AMSN) and the MIST research program of the Dutch Ministry of Defense. The INS submission was partly supported by the MIME project of the creative industries knowledge and innovation network CLICKNL., peer-reviewed