PaperSummary10 : Applying eigencontours to polar mask-based instance segmentation
The paper discusses enhancing instance segmentation, which identifies object boundaries in images using eigencontours. Eigencontours are data driven contour descriptors derived from singular value decomposition (SVD). The paper incorporates eigencontours into the PolarMask framework that improves segmentation accuracy and efficiency over pixelwise classification or other contour-based methods.
The key steps of the method are:
- Construction of Eigencontours: First, object boundaries are converted to star-convex using polar coordinates. Second, SVD is performed on the contour matrix to derive eigencontours. And then, a rank-M approximation is used to represent boundaries in reduced dimensional space.
- Integration with PolarMask: Radial coordinate regression are replaced with eigencontour coefficient regression. A network is designed using an encoder-decoder architecture with ResNet50 as the backbone. Output coefficients are used to reconstruct object boundaries in eigencontour space.
- Postprocessing: Confidence scoring, non-maximum suppression and contour reconstruction is applied to generate segmentation masks.
The proposed method outperforms the original PolarMask, demonstrating the potential of eigencontours for efficient and accurate instance segmentation. It provides efficient representation which reduces computational requirements and performs better in describing curved and complex boundaries.
References: