BoundaryNet: An Attentive Deep Network with Fast Marching Distance Maps for Semi-automatic Layout Annotation

Precise boundary annotations of image regions can be crucial for downstream\napplications which rely on region-class semantics. Some document collections\ncontain densely laid out, highly irregular and overlapping multi-class region\ninstances with large range in aspect ratio. Fully automatic boundary estimation\napproaches tend to be data intensive, cannot handle variable-sized images and\nproduce sub-optimal results for aforementioned images. To address these issues,\nwe propose BoundaryNet, a novel resizing-free approach for high-precision\nsemi-automatic layout annotation. The variable-sized user selected region of\ninterest is first processed by an attention-guided skip network. The network\noptimization is guided via Fast Marching distance maps to obtain a good quality\ninitial boundary estimate and an associated feature representation. These\noutputs are processed by a Residual Graph Convolution Network optimized using\nHausdorff loss to obtain the final region boundary. Results on a challenging\nimage manuscript dataset demonstrate that BoundaryNet outperforms strong\nbaselines and produces high-quality semantic region boundaries. Qualitatively,\nour approach generalizes across multiple document image datasets containing\ndifferent script systems and layouts, all without additional fine-tuning. We\nintegrate BoundaryNet into a document annotation system and show that it\nprovides high annotation throughput compared to manual and fully automatic\nalternatives.\n