SkeletonNetV2: A Dense Channel Attention Blocks for Skeleton Extraction

Sabari Nathan, Priya Kansal; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2142-2149

Abstract


Geometrical analysis of a shape through skeletonization has some of very important high- and low-level application which includes tracking, manipulation, retrieval, representation, registration, recognition, and compression. The task of skeletonization is defined as the generation of the medial axis of the shape while preserving its original topology and geometry. While the earlier approaches are mainly based on the extracting the skeleton and then pruning the unwanted branches, the present study proposes a novel convolutional neural network based method to perform this task. The proposed architecture is an encoder-decoder network which leverage the benefits of coordinated convolutional layer and multi-level supervision to prevent the loss of information between the extracted skeleton and the ground truth. The dense attention block is used as the backbone blocks in encoder and decoder block. This architecture is performing better than the state of art on not only skeletonization of image task but also skeletonization from the point cloud. This method achieved a F1 score of 0.7961 on Pixel Skeleton dataset and a Chamfer Distance (CD) score of 1.9561 on Point skeleton dataset.

Related Material


[pdf]
[bibtex]
@InProceedings{Nathan_2021_ICCV, author = {Nathan, Sabari and Kansal, Priya}, title = {SkeletonNetV2: A Dense Channel Attention Blocks for Skeleton Extraction}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, month = {October}, year = {2021}, pages = {2142-2149} }