Conceptos Básicos
AdaContour employs multiple local contour representations to effectively capture complex and irregular object shapes, outperforming existing angle-based contour descriptors that use a single global representation.
Resumen
The paper proposes AdaContour, an adaptive contour descriptor that uses a hierarchical representation to accurately model complex object shapes. Existing angle-based contour descriptors, such as Eigen and ESE, suffer from lossy representations for non-starconvex shapes due to their use of a single global inner center and polar coordinate parameterization.
To address this, the paper introduces a novel hierarchical encoding procedure that recursively subdivides the shape until sufficiently regular regions are identified. Each refined region is then encoded by a local contour, allowing AdaContour to better capture irregular boundaries. The paper also utilizes robust subspace recovery techniques and basis-sharing mechanisms to efficiently represent the local contours.
Experiments show that AdaContour outperforms existing angle-based descriptors in accurately representing complex shapes, while retaining computational efficiency. The paper also demonstrates the integration of AdaContour into an object detection framework for instance segmentation, exhibiting favorable performance compared to conventional methods.
The key highlights of the paper are:
Hierarchical encoding of object shapes using multiple local contour representations to handle non-starconvex shapes.
Robust subspace recovery and basis-sharing techniques to enable efficient and accurate shape approximation.
Validation of AdaContour's effectiveness through instance segmentation experiments, outperforming existing angle-based descriptors.
Estadísticas
The paper reports the following key metrics:
IOU (Intersection over Union) between ground-truth and reconstructed object masks on the KINS, SBD, and COCO2017 datasets.
Average Precision (AP) scores for instance segmentation on the SBD validation dataset.