Citation Context

... for other categories. Other methods aim to cope with different data distributions in the source and target domains for the same categories, e.g. due to lighting, background and view-point variations =-=[13, 23, 41]-=-. These and other related methods adapt classifiers or kernels while using standard image features. Differently to this work, we here transfer image representations trained on the source task. More si...

Citation Context

...rm can refer to two completely different concepts (see Figure 5). To alleviate these concerns, it might be possible to tune our model to account for dataset biases and thereby improve its performance =-=[20, 47]-=-. Tuning biases in a completely unsupervised setting is also an interesting research direction. Testing our model involves convolving each test image with an exhaustive model of around 250 components....

Citation Context

...ich trained the shared part bases to accelerate the detection. Note that [34] learned a shared filter bases, while our model learns a shared classifier, which result in a quite different formulation. =-=[26]-=- also proposed a multi-task model to handle dataset bias. The multi-task idea in this paper is motivated by works on face recognition across different domains, such as [28, 5]. Context has been used i...

Citation Context

...tiple modalities or multiple views of the same data are available. This notion was also exploited in the context of Multi-Task learning in [23]. Recently the dataset bias problem has been explored in =-=[26]-=-. The proposed approach is based on the combination of a specific and a common discriminative model across several tasks, following the same idea of the original multi-task SVM [18]. The novelty is in...

Citation Context

...ar feature representations. 1. Introduction It has become increasingly clear that there is a significant bias between available labeled visual training data and the data encountered in the real world =-=[28, 16]-=-. Unfortunately, supervised classifiers trained on one distribution often fail when faced with a different distribution at test time. Domain adaptation techniques offer a way to transfer information l...

Citation Context

...ces surprisingly accurate reconstructions, even when the database does not contain the category of the HOG template. We note that this method may be subject to dataset bias [25] but could be overcome =-=[15]-=-. We also point out that a similar nearest neighbor method is used in brain research to visualize what a person might be seeing [20]. 3.2. Algorithm B: Ridge Regression Unfortunately, running an objec...

Citation Context

...y) . In both cases, the techniques proposed so far to rectify the distribution mismatch, regardless of its underlying reason, range from max-margin classifiers able to adapt their learning parameters =-=[15]-=- to methods attempting to learn how to project the data in a new intermediate space, where the features lose the specific bias [13]. But what if broadly adopted features and classifiers would be part ...

Citation Context

...ds is that they have been demonstrated only on a handful of object categories, namely, car [14, 23], bicycle [14] and motorcycle[15]. We show the feasibility of our approach for a varied set of 20 object categories. Domain Adaptation Domain adaptation is the standard approach to alleviate dataset bias caused by a difference in the statistical distributions between training and test data. In computer vision, several domain adaptation models have been proposed for object categorization. These can be divided into supervised methods that use a small number of labeled examples in the target domain [11, 12, 13], and unsupervised methods that use only unlabeled target examples [8, 9]. For detection tasks, a class of supervised models that adjust the parameters of a classifier trained on source data have been proposed for linear support vector machines [1, 28], as well as for variants of SVMs, e.g., adaptive latent SVM [27] or adaptive exemplar SVM [2]. 4 BMVC 2014 BAOCHEN SUN, KATE SAENKO: FROM VIRTUAL TO REALITY Figure 2: Some of the 3D models used in this paper. Recently there has been considerable interest in scalable training of many object detectors. The supervised domain adaptation method of [7...

Citation Context

...re workers on AMT may not originate from the locations being tested i.e. the worker might be from Paris while the task images are from Boston. This can also be thought of as a problem of dataset bias =-=[23, 37]-=- i.e., a sampling bias. Despite significant differences in the cities, we find that the features are able to generalize reasonably well across cities. Surprisingly, for four of the eight cities, train...

Citation Context

...d appearances with a large degree of occlusion and truncation. Transform-based domain adaptation overcomes the bias by learning a transformation between datasets. In contrast to classifier adaptation =-=[1, 22, 3, 11]-=-, learning a transformation between feature spaces directly allows us to perform adaptation even for (new) categories that are not present in both datasets. Especially for large-scale recognition with...