| Abstract: |
| To address imbalanced object sizes in image segmentation, this work formulates the problem via optimal transport theory and a Laguerre cell decomposition. The dual variable of the volume constraint is interpreted as a learnable bias, and an iterative network-embedding laye (VP-Sparsemax) is proposed to solve the smooth semi-dual formulation while incorporating spatial information. Compared to softmax, VP-Sparsemax improves volume preservation after argmax due to its sparsity. Experiments on four dataset across three segmentation baselines demonstrate superior performance, especially for small targets that are easily overlooked. |
|