Fig 1.
Coarse semantic segmentation results of the PASCAL VOC dataset based on the FCN and DeepLab-CRF model.
Different colors represent different classes. (a) Input image (b) Segmentation results from FCN (first two rows) and DeepLab-CRF (last two rows) (c) Ground truth.
Fig 2.
Illustration of deconvolution operations.
(a) Convolution (many-to-one) (b) Deconvolution (one-to-many).
Fig 3.
Superpixel segmentation results from the GS method.
Different colors represent different superpixels. (a) Input image (b) Superpixels from GS method (c) Semantic labels (d) Superpixels from GS method with semantic labels.
Fig 4.
Examples that our method based on GS superpixels produced better results than the FCN model.
Different colors represent different classes. (a) Input image (b) Segmentation results from FCN (c) Segmentation results from FCN-GS (d) FCN-GS-v2 (e) Ground truth.
Fig 5.
Examples that our method based on GS superpixels produced better results than the DeepLab-CRF model.
Different colors represent different classes. (a) Input image (b) Segmentation results from DeepLab-CRF (c) Segmentation results from DeepLab-CRF-GS (d) Segmentation results from DeepLab-CRF-GS-v2 (e) Ground truth.
Fig 6.
Examples that our method based on DBSCAN superpixels produced better results than the FCN model.
Different colors represent different classes. (a) Input image (b) Segmentation results from FCN (c) Segmentation results from FCN-DBSCAN (d) FCN-DBSCAN-v2 (e) Ground truth.
Table 1.
Performance of our proposed models on the PASCAL VOC 2011 and 2012 test sets compared to other state-of-art methods.
Table 2.
Evaluation results of the PASCAL VOC 2012 test set.