Multiple instance learning: A survey of problem characteristics and applications - ScienceDirect

Fig. 2. Illustration of two decisions boundaries on a fictive problem. While only the purple boundary correctly classifies all instances, both of them achieve perfect bag classification. This is because, in that case, false positive and false negative instances do not impact on bag labels. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 3. Illustration of intra-bag similarity between instances: the patches are overlapping, and thus, share similarities with each other. Fig. 4. Example of co-occurrence and similarity between instances: three segments contain grass and forest and are therefore very similar. Moreover, since this is an image of a bear, the background is more likely to be nature than a nuclear central control room. Fig. 5. For the same concept ants, there can be many data clusters (modes) in feature space corresponding to different poses, colors and castes. (For interpretation of the refer

Fig. 2. Illustration of two decisions boundaries on a fictive problem. While only the purple boundary correctly classifies all instances, both of them achieve perfect bag classification. This is because, in that case, false positive and false negative instances do not impact on bag labels. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 3. Illustration of intra-bag similarity between instances: the patches are overlapping, and thus, share similarities with each other. Fig. 4. Example of co-occurrence and sim