Command Palette
Search for a command to run...
Chen Liang-Chieh Papandreou George Schroff Florian Adam Hartwig

摘要
在本工作中,我们重新审视了空洞卷积(atrous convolution)这一强大工具在语义图像分割任务中的应用。空洞卷积能够显式调整卷积核的感受野,同时控制深度卷积神经网络所生成特征响应的分辨率。为应对多尺度目标分割的挑战,我们设计了若干模块,通过级联或并行方式使用空洞卷积,并采用多个空洞率来捕捉多尺度上下文信息。此外,我们提出对先前提出的空洞空间金字塔池化(Atrous Spatial Pyramid Pooling)模块进行增强,引入编码全局上下文信息的图像级特征,进一步提升模型性能。我们还详细阐述了实现细节,并分享了训练该系统过程中积累的经验。所提出的“DeepLabv3”系统在无需使用DenseCRF后处理的情况下,显著优于我们之前的DeepLab版本,并在PASCAL VOC 2012语义图像分割基准测试中达到了与当前其他先进模型相当的性能水平。
代码仓库
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| dichotomous-image-segmentation-on-dis-te1 | DeeplabV3+ | E-measure: 0.772 HCE: 234 MAE: 0.102 S-Measure: 0.694 max F-Measure: 0.601 weighted F-measure: 0.506 |
| dichotomous-image-segmentation-on-dis-te2 | DeeplabV3+ | E-measure: 0.813 HCE: 516 MAE: 0.105 S-Measure: 0.729 max F-Measure: 0.681 weighted F-measure: 0.587 |
| dichotomous-image-segmentation-on-dis-te3 | DeeplabV3+ | E-measure: 0.833 HCE: 999 MAE: 0.102 S-Measure: 0.749 max F-Measure: 0.717 weighted F-measure: 0.623 |
| dichotomous-image-segmentation-on-dis-te4 | DeeplabV3+ | E-measure: 0.820 HCE: 3709 MAE: 0.111 S-Measure: 0.744 max F-Measure: 0.715 weighted F-measure: 0.621 |
| dichotomous-image-segmentation-on-dis-vd | DeeplabV3+ | E-measure: 0.796 HCE: 1520 MAE: 0.114 S-Measure: 0.716 max F-Measure: 0.660 weighted F-measure: 0.568 |
| semantic-segmentation-on-cityscapes | DeepLabv3 (ResNet-101, coarse) | Mean IoU (class): 81.3% |
| semantic-segmentation-on-cityscapes-val | DeepLabv3 (Dilated-ResNet-101) | mIoU: 78.5% |
| semantic-segmentation-on-pascal-voc-2012 | DeepLabv3-JFT | Mean IoU: 86.9% |
| semantic-segmentation-on-pascal-voc-2012-val | DeepLabv3-JFT | mIoU: 82.7% |
| semantic-segmentation-on-selma | DeepLabV3 | mIoU: 70.7 |