Dynamic Snake Convolution

[yanrihong]

Describe the feature

Motivation The motivation behind this feature request is to integrate the recently proposed Dynamic Snake Convolution (DSCNet) model for tubular structure segmentation into the mmsegmentation library. The DSCNet model has demonstrated significant improvements in accuracy and continuity for the segmentation of elongated tubular structures. Given the importance of tubular structures in various domains such as medical imaging and road network extraction, incorporating DSCNet into mmsegmentation would enhance the library's capabilities in handling such specialized segmentation tasks.

Related resources The official code release for the Dynamic Snake Convolution (DSCNet) model can be found on GitHub: DSCNet GitHub Repository. This implementation is based on PyTorch and follows the architecture presented in the paper "Dynamic Snake Convolution Based on Topological Geometric Constraints for Tubular Structure Segmentation" by Qi et al., published at ICCV 2023.

Additional context The DSCNet model introduces a specialized network structure tailored to the characteristics of tubular structures, allowing the convolutional kernel to dynamically adapt its shape for effective feature learning while ensuring it stays focused on the target structure. The integration of DSCNet into mmsegmentation would provide users with a powerful tool for accurate and continuous segmentation of tubular structures in 2D and 3D data.

[hqzqaq]

Have you successfully configured DSCNet? I tried, but I didn't succeed because I am not familiar with the configuration of mmsegmentation.

[yanrihong]

Have you successfully configured DSCNet? I tried, but I didn't succeed because I am not familiar with the configuration of mmsegmentation.

not yet 。。。。