XiongXinyu ; WuZihuang ; TanShuangyi ; LiWenxue ; TangFeilong ; ChenYing ; LiSiying ; MaJie ; LiGuanbin

摘要
图像分割在视觉理解中发挥着重要作用。近年来,新兴的视觉基础模型在各种任务上不断取得优异的性能。在此背景下,本文证明了 Segment Anything Model 2 (SAM2) 可以作为 U 形分割模型的强大编码器。我们提出了一种简单而有效的框架,称为 SAM2-UNet,用于多功能图像分割。具体而言,SAM2-UNet 采用了 SAM2 的 Hiera 主干作为编码器,而解码器则使用经典的 U 形设计。此外,编码器中插入了适配器,以实现参数高效的微调。初步实验表明,在各种下游任务(如伪装物体检测、显著物体检测、海洋动物分割、镜面检测和息肉分割)中,我们的 SAM2-UNet 能够轻松超越现有的专门化最先进方法,无需复杂的附加技术。项目页面:\url{https://github.com/WZH0120/SAM2-UNet}。
代码仓库
wzh0120/sam2-unet
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| image-segmentation-on-mas3k | SAM2-UNet | E-measure: 0.943 MAE: 0.021 S-measure: 0.903 mIoU: 0.799 |
| image-segmentation-on-msd-mirror-segmentation | SAM2-UNet | F-measure: 0.957 IoU: 0.918 MAE: 0.022 |
| image-segmentation-on-pmd | SAM2-UNet | F-measure: 0.826 IoU: 0.728 MAE: 0.027 |
| image-segmentation-on-rmas | SAM2-UNet | E-measure: 0.944 MAE: 0.022 S-measure: 0.874 mIoU: 0.738 |
| salient-object-detection-on-dut-omron-2 | SAM2-UNet | E-measure: 0.912 MAE: 0.039 S-measure: 0.884 |
| salient-object-detection-on-duts-te-1 | SAM2-UNet | E-measure: 0.959 MAE: 0.020 Smeasure: 0.934 |
| salient-object-detection-on-ecssd-1 | SAM2-UNet | E-measure: 0.970 MAE: 0.020 S-measure: 0.950 |
| salient-object-detection-on-hku-is-1 | SAM2-UNet | E-measure: 0.971 MAE: 0.019 S-measure: 0.941 |
| salient-object-detection-on-pascal-s-1 | SAM2-UNet | E-measure: 0.931 MAE: 0.043 S-measure: 0.894 |