PaperSummary08 : PolarMask
The paper introduces, PolarMask, an anchor box-free, single shot instance segmentation method. It uses polar coordinates to predict instance masks via instance center classification and dense distance regression.
The key points are:
- Polar Representation: Instance masks are represented as contours, modeled using a central point and rays extending to the contour at defined angles. This simplifies mask generation and assembly.
- Polar Centerness: A weighting mechanism to prioritize high-quality center points during training and inference, improving mask precision.
- Polar IoU Loss: A loss function introduced for dense distance regression that enhances the prediction accuracy of contours.
It is simple comparable to single-shot detection method with negligible computational head. It easily integrates into off-the-shelf detectors like FCOS or RetinaNet. It is faster during inference compared to other methods such as TensorMask. However, it faces challenges with corner cases such as complex contours or rays that miss the contour entirely. And mask quality depends on the number of rays with diminishing returns for excessive rays.
Overall, PolarMask provides an effective framework for instance segmentation, changing the focus of contours from pixel-wise classification.
References:
