
摘要
我们提出了首个利用不确定性进行RGB-D显著性检测的随机框架,该框架通过从数据标注过程中学习来实现。现有的RGB-D显著性检测模型将此任务视为一个点估计问题,即通过确定性的学习流程预测单一的显著性图。我们认为,然而,这种确定性的解决方案相对不够明确。受显著性数据标注过程的启发,我们提出了一种生成架构以实现概率性的RGB-D显著性检测,该架构利用潜在变量来建模标注变化。我们的框架包含两个主要模型:1)生成模型,它将输入图像和潜在变量映射到随机的显著性预测;2)推理模型,它通过从真实或近似后验分布中采样潜在变量来逐步更新潜在变量。生成模型是一个编码器-解码器显著性网络。为了推断潜在变量,我们引入了两种不同的解决方案:i)条件变分自编码器(Conditional Variational Auto-encoder),其中包含一个额外的编码器来近似潜在变量的后验分布;ii)交替反向传播技术(Alternating Back-Propagation),该技术直接从真实的后验分布中采样潜在变量。在六个具有挑战性的RGB-D基准数据集上的定性和定量结果表明,我们的方法在学习显著性图的分布方面表现出优越性能。源代码已通过我们的项目页面公开提供:https://github.com/JingZhang617/UCNet。
代码仓库
JingZhang617/UCNet
官方
pytorch
GitHub 中提及
DengPingFan/SOC-DataAug
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| rgb-d-salient-object-detection-on-des | UCNet-ABP | Average MAE: 0.016 S-Measure: 94.0 |
| rgb-d-salient-object-detection-on-des | UCNet-CVAE | Average MAE: 0.016 S-Measure: 93.7 |
| rgb-d-salient-object-detection-on-lfsd | UCNet-ABP | Average MAE: 0.065 S-Measure: 86.6 |
| rgb-d-salient-object-detection-on-lfsd | UCNet-CVAE | Average MAE: 0.065 S-Measure: 86.8 |
| rgb-d-salient-object-detection-on-nju2k | UCNet-CVAE | Average MAE: 0.039 S-Measure: 90.2 |
| rgb-d-salient-object-detection-on-nju2k | UCNet-ABP | Average MAE: 0.039 S-Measure: 90.0 |
| rgb-d-salient-object-detection-on-nlpr | UCNet-CAVE | Average MAE: 0.025 S-Measure: 91.7 |
| rgb-d-salient-object-detection-on-nlpr | UCNet-ABP | Average MAE: 0.024 S-Measure: 91.9 |
| rgb-d-salient-object-detection-on-sip | UCNet-CVAE | Average MAE: 0.045 S-Measure: 88.3 |
| rgb-d-salient-object-detection-on-sip | UCNet-ABP | Average MAE: 0.049 S-Measure: 87.6 |
| rgb-d-salient-object-detection-on-stere | UCNet-CVAE | Average MAE: 0.039 S-Measure: 89.8 |
| rgb-d-salient-object-detection-on-stere | UCNet-ABP | Average MAE: 0.037 S-Measure: 90.4 |
| salient-object-detection-on-dut-omron | UCNet-ABP | MAE: 0.050 S-Measure: 0.843 |
| salient-object-detection-on-dut-omron | UCNet-CVAE | MAE: 0.051 S-Measure: 0.839 |
| salient-object-detection-on-duts-te | UCNet-CVAE | MAE: 0.034 S-Measure: 0.888 mean E-Measure: 0.927 mean F-Measure: 0.860 |
| salient-object-detection-on-duts-te | UCNet-ABP | MAE: 0.034 S-Measure: 0.890 mean E-Measure: 0.931 mean F-Measure: 0.864 |
| salient-object-detection-on-ecssd | UCNet-CVAE | MAE: 0.035 S-Measure: 0.921 |
| salient-object-detection-on-hku-is | UCNet-CVAE | MAE: 0.026 S-Measure: 0.921 |
| salient-object-detection-on-hku-is | UCNet-ABP | MAE: 0.027 S-Measure: 0.917 |
| salient-object-detection-on-soc | UCNet-APB | Average MAE: 0.091 S-Measure: 0.842 mean E-Measure: 0.868 |
| salient-object-detection-on-soc | UCNet-CVAE | Average MAE: 0.089 S-Measure: 0.849 mean E-Measure: 0.872 |