摘要
少样本语义分割旨在仅利用目标类别少量标注的支持图像，从查询图像中实现对目标物体的分割。这一具有挑战性的任务要求模型能够理解多层次的视觉线索，并精确分析查询图像与支持图像之间的细粒度对应关系。为应对该问题，本文提出超相关性压缩网络（Hypercorrelation Squeeze Networks, HSNet），该方法融合多层级特征相关性与高效的四维卷积操作。HSNet从不同深度的卷积中间层提取多样化特征，并构建一组四维相关性张量，即超相关性（hypercorrelations）。在金字塔式架构中，通过高效的中心枢轴四维卷积，该方法以由粗到细的方式逐步压缩超相关性中的高层语义信息与低层几何结构信息，最终生成精确的分割掩码。在PASCAL-5i、COCO-20i和FSS-1000等标准少样本分割基准上的显著性能提升，充分验证了所提方法的有效性。
代码仓库
juhongm999/hsnet
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| few-shot-semantic-segmentation-on-coco-20i-1 | HSNet (ResNet-50) | FB-IoU: 68.2 Mean IoU: 39.2 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-coco-20i-1 | HSNet (ResNet-101) | FB-IoU: 69.1 Mean IoU: 41.2 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-coco-20i-5 | HSNet (ResNet-101) | FB-IoU: 72.4 Mean IoU: 49.5 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-coco-20i-5 | HSNet (ResNet-50) | FB-IoU: 70.7 Mean IoU: 46.9 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-fss-1000-1 | HSNet (VGG-16) | Mean IoU: 82.3 |
| few-shot-semantic-segmentation-on-fss-1000-1 | HSNet (ResNet-101) | Mean IoU: 86.5 |
| few-shot-semantic-segmentation-on-fss-1000-1 | HSNet (ResNet-50) | Mean IoU: 85.5 |
| few-shot-semantic-segmentation-on-fss-1000-5 | HSNet (ResNet-50) | Mean IoU: 87.8 |
| few-shot-semantic-segmentation-on-fss-1000-5 | HSNet (VGG-16) | Mean IoU: 85.8 |
| few-shot-semantic-segmentation-on-fss-1000-5 | HSNet (ResNet-101) | Mean IoU: 88.5 |
| few-shot-semantic-segmentation-on-pascal-5i-1 | HSNet (ResNet-50) | FB-IoU: 76.7 Mean IoU: 64.0 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-pascal-5i-1 | HSNet (ResNet-101) | FB-IoU: 77.6 Mean IoU: 66.2 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-pascal-5i-1 | HSNet (VGG-16) | FB-IoU: 73.4 Mean IoU: 59.7 |
| few-shot-semantic-segmentation-on-pascal-5i-5 | HSNet (ResNet-101) | FB-IoU: 80.6 Mean IoU: 70.4 learnable parameters (million): 2.5 |
| few-shot-semantic-segmentation-on-pascal-5i-5 | HSNet (VGG-16) | FB-IoU: 76.6 Mean IoU: 64.1 |
| few-shot-semantic-segmentation-on-pascal-5i-5 | HSNet (ResNet-50) | FB-IoU: 80.6 Mean IoU: 69.5 learnable parameters (million): 2.5 |