UMBC RosSeg

The In-House dataset is acquired using a Husarion ROSbot 2.0 and ROSbot 2.0 Pro with the collection speed set to 5 frames per second from a suburban university environment. We may split the information into subgroups for noon, dusk, and dawn in order to depict our surroundings under various lighting situations. We have assembled 17 sequences totaling 8080 frames, of which 1619 have been manually analyzed using an open-source pixel annotation program. Since nearby photographs are highly similar to one another, we decide to annotate every five images. Since the annotation procedure may be highly time-consuming, we employ soft-labeling while annotating the In-House dataset, which enables us to proceed through the frames more quickly. The annotation method we employ enables us to create minute annotations inside an image's objects, and the categorization would encompass related pixels. This approach may result in less-than-perfect annotations and some performance accuracy loss, but the loss would not be great enough to justify producing pixel-perfect annotations.