Qianxiong Xu, Xuanyi Liu, Lanyun Zhu, Guosheng Lin, Cheng Long, Ziyue Li, Rui Zhao
Many few-shot segmentation (FSS) methods use cross attention to fuse support foreground (FG) into query features, regardless of the quadratic complexity. A recent advance Mamba can also well capture intra-sequence dependencies, yet the complexity is only linear. Hence, we aim to devise a cross (attention-like) Mamba to capture inter-sequence dependencies for FSS. A simple idea is to scan on support features to selectively compress them into the hidden state, which is then used as the initial hidden state to sequentially scan query features. Nevertheless, it suffers from (1) support forgetting issue: query features will also gradually be compressed when scanning on them, so the support features in hidden state keep reducing, and many query pixels cannot fuse sufficient support features; (2) intra-class gap issue: query FG is essentially more similar to itself rather than to support FG, i.e., query may prefer not to fuse support features but their own ones from the hidden state, yet the success of FSS relies on the effective use of support information. To tackle them, we design a hybrid Mamba network (HMNet), including (1) a support recapped Mamba to periodically recap the support features when scanning query, so the hidden state can always contain rich support information; (2) a query intercepted Mamba to forbid the mutual interactions among query pixels, and encourage them to fuse more support features from the hidden state. Consequently, the support information is better utilized, leading to better performance. Extensive experiments have been conducted on two public benchmarks, showing the superiority of HMNet. The code is available at https://github.com/Sam1224/HMNet.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Few-Shot Learning | COCO-20i (5-shot) | FB-IoU | 77.6 | HMNet (ResNet-50) |
| Few-Shot Learning | COCO-20i (5-shot) | Mean IoU | 58.9 | HMNet (ResNet-50) |
| Few-Shot Learning | COCO-20i (5-shot) | FB-IoU | 75.5 | HMNet (VGG-16) |
| Few-Shot Learning | COCO-20i (5-shot) | Mean IoU | 54.5 | HMNet (VGG-16) |
| Few-Shot Learning | PASCAL-5i (1-Shot) | FB-IoU | 81.6 | HMNet (ResNet-50) |
| Few-Shot Learning | PASCAL-5i (1-Shot) | Mean IoU | 70.4 | HMNet (ResNet-50) |
| Few-Shot Learning | PASCAL-5i (1-Shot) | FB-IoU | 79.2 | HMNet (VGG-16) |
| Few-Shot Learning | PASCAL-5i (1-Shot) | Mean IoU | 67.3 | HMNet (VGG-16) |
| Few-Shot Learning | COCO-20i (1-shot) | FB-IoU | 74.5 | HMNet (ResNet-50) |
| Few-Shot Learning | COCO-20i (1-shot) | Mean IoU | 52.1 | HMNet (ResNet-50) |
| Few-Shot Learning | COCO-20i (1-shot) | FB-IoU | 72.6 | HMNet (VGG-16) |
| Few-Shot Learning | COCO-20i (1-shot) | Mean IoU | 49.1 | HMNet (VGG-16) |
| Few-Shot Learning | PASCAL-5i (5-Shot) | FB-IoU | 84.4 | HMNet (ResNet-50) |
| Few-Shot Learning | PASCAL-5i (5-Shot) | Mean IoU | 74.1 | HMNet (ResNet-50) |
| Few-Shot Learning | PASCAL-5i (5-Shot) | FB-IoU | 82.6 | HMNet (VGG-16) |
| Few-Shot Learning | PASCAL-5i (5-Shot) | Mean IoU | 71.1 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | COCO-20i (5-shot) | FB-IoU | 77.6 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | COCO-20i (5-shot) | Mean IoU | 58.9 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | COCO-20i (5-shot) | FB-IoU | 75.5 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | COCO-20i (5-shot) | Mean IoU | 54.5 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | PASCAL-5i (1-Shot) | FB-IoU | 81.6 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | PASCAL-5i (1-Shot) | Mean IoU | 70.4 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | PASCAL-5i (1-Shot) | FB-IoU | 79.2 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | PASCAL-5i (1-Shot) | Mean IoU | 67.3 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | COCO-20i (1-shot) | FB-IoU | 74.5 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | COCO-20i (1-shot) | Mean IoU | 52.1 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | COCO-20i (1-shot) | FB-IoU | 72.6 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | COCO-20i (1-shot) | Mean IoU | 49.1 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | PASCAL-5i (5-Shot) | FB-IoU | 84.4 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | PASCAL-5i (5-Shot) | Mean IoU | 74.1 | HMNet (ResNet-50) |
| Few-Shot Semantic Segmentation | PASCAL-5i (5-Shot) | FB-IoU | 82.6 | HMNet (VGG-16) |
| Few-Shot Semantic Segmentation | PASCAL-5i (5-Shot) | Mean IoU | 71.1 | HMNet (VGG-16) |
| Meta-Learning | COCO-20i (5-shot) | FB-IoU | 77.6 | HMNet (ResNet-50) |
| Meta-Learning | COCO-20i (5-shot) | Mean IoU | 58.9 | HMNet (ResNet-50) |
| Meta-Learning | COCO-20i (5-shot) | FB-IoU | 75.5 | HMNet (VGG-16) |
| Meta-Learning | COCO-20i (5-shot) | Mean IoU | 54.5 | HMNet (VGG-16) |
| Meta-Learning | PASCAL-5i (1-Shot) | FB-IoU | 81.6 | HMNet (ResNet-50) |
| Meta-Learning | PASCAL-5i (1-Shot) | Mean IoU | 70.4 | HMNet (ResNet-50) |
| Meta-Learning | PASCAL-5i (1-Shot) | FB-IoU | 79.2 | HMNet (VGG-16) |
| Meta-Learning | PASCAL-5i (1-Shot) | Mean IoU | 67.3 | HMNet (VGG-16) |
| Meta-Learning | COCO-20i (1-shot) | FB-IoU | 74.5 | HMNet (ResNet-50) |
| Meta-Learning | COCO-20i (1-shot) | Mean IoU | 52.1 | HMNet (ResNet-50) |
| Meta-Learning | COCO-20i (1-shot) | FB-IoU | 72.6 | HMNet (VGG-16) |
| Meta-Learning | COCO-20i (1-shot) | Mean IoU | 49.1 | HMNet (VGG-16) |
| Meta-Learning | PASCAL-5i (5-Shot) | FB-IoU | 84.4 | HMNet (ResNet-50) |
| Meta-Learning | PASCAL-5i (5-Shot) | Mean IoU | 74.1 | HMNet (ResNet-50) |
| Meta-Learning | PASCAL-5i (5-Shot) | FB-IoU | 82.6 | HMNet (VGG-16) |
| Meta-Learning | PASCAL-5i (5-Shot) | Mean IoU | 71.1 | HMNet (VGG-16) |