Inference with model uncertainty on indoor scene for semantic segmentation

This paper presents a novel semantic segmentation method using model uncertainty on indoor scene. Recently, deep convolutional encoder-decoder neural networks achieved good segmentation performance and we also implement this network. Toward practical use of the semantic segmentation, it is necessary to recognize accurately whether the object is known or not. However, many architectures do not consider an uncertainty of the segmentation result. Bayesian SegNet enables it to produce an uncertainty of the segmentation results with a measure of model uncertainty but the uncertainty is not used for segmentation itself. Our study aims the improvement of classification accuracy for semantic segmentation in indoor scene understanding using model uncertainty. Therefore, we use SUN RGB-D indoor scene dataset for the training and testing of our network. Our method rejects the uncertain region and classifies it as an unknown object using the model uncertainty. As a result, we succeeded in reducing wrong predictions. Furthermore, our method showed good classification performance in the comparison with other architectures such as FCN, SegNet, Bayesian SegNet, and so on. We achieved improvement of classification accuracy by our method as shown in the evaluation results.