eduzhai > Applied Sciences > Engineering >

Fully Convolutional Open Set Segmentation

  • Save

... pages left unread,continue reading

Document pages: 34 pages

Abstract: In semantic segmentation knowing about all existing classes is essential toyield effective results with the majority of existing approaches. However,these methods trained in a Closed Set of classes fail when new classes arefound in the test phase. It means that they are not suitable for Open Setscenarios, which are very common in real-world computer vision and remotesensing applications. In this paper, we discuss the limitations of Closed Setsegmentation and propose two fully convolutional approaches to effectivelyaddress Open Set semantic segmentation: OpenFCN and OpenPCS. OpenFCN is basedon the well-known OpenMax algorithm, configuring a new application of thisapproach in segmentation settings. OpenPCS is a fully novel approach based onfeature-space from DNN activations that serve as features for computing PCA andmulti-variate gaussian likelihood in a lower dimensional space. Experimentswere conducted on the well-known Vaihingen and Potsdam segmentation datasets.OpenFCN showed little-to-no improvement when compared to the simpler and muchmore time efficient SoftMax thresholding, while being between some orders ofmagnitude slower. OpenPCS achieved promising results in almost all experimentsby overcoming both OpenFCN and SoftMax thresholding. OpenPCS is also areasonable compromise between the runtime performances of the extremely fastSoftMax thresholding and the extremely slow OpenFCN, being close able to runclose to real-time. Experiments also indicate that OpenPCS is effective, robustand suitable for Open Set segmentation, being able to improve the recognitionof unknown class pixels without reducing the accuracy on the known classpixels.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...