Class-Agnostic Segmentation Loss and Its Application to Salient Object Detection and Segmentation

In this paper we present a novel loss function, called class-agnostic\nsegmentation (CAS) loss. With CAS loss the class descriptors are learned during\ntraining of the network. We don't require to define the label of a class\na-priori, rather the CAS loss clusters regions with similar appearance together\nin a weakly-supervised manner. Furthermore, we show that the CAS loss function\nis sparse, bounded, and robust to class-imbalance. We first apply our CAS loss\nfunction with fully-convolutional ResNet101 and DeepLab-v3 architectures to the\nbinary segmentation problem of salient object detection. We investigate the\nperformance against the state-of-the-art methods in two settings of low and\nhigh-fidelity training data on seven salient object detection datasets. For\nlow-fidelity training data (incorrect class label) class-agnostic segmentation\nloss outperforms the state-of-the-art methods on salient object detection\ndatasets by staggering margins of around 50%. For high-fidelity training data\n(correct class labels) class-agnostic segmentation models perform as good as\nthe state-of-the-art approaches while beating the state-of-the-art methods on\nmost datasets. In order to show the utility of the loss function across\ndifferent domains we then also test on general segmentation dataset, where\nclass-agnostic segmentation loss outperforms competing losses by huge margins.\n