Fig 1.
Key technologies of the proposed method.
Fig 2.
TDCN structure.
Table 1.
Important notations and the corresponding definitions in this work.
Fig 3.
Edge operators, normal convolutions, and trainable absolute gradient (TAG) layer.
Fig 4.
Structure of difference convolution module.
Fig 5.
Typical self-attention layer and proposed boundary-aware attention layer.
Fig 6.
Transformer with boundary-aware attention.
Fig 7.
Structure of the head function.
Fig 8.
Precision-recall curves on BSDS dataset.
Table 2.
Performances on BSDS.
Table 3.
Performances on NYUD.
Fig 9.
Number of parameters and corresponding performances (ODS) of different structures.
Fig 10.
Prediction results for BSDS (top), NYUD-RGB (middle), and NYUD-HHA (bottom).
Table 4.
Performance comparison between unified prediction and dataset-specific prediction.
Fig 11.
Prediction results under complete TDCN and corresponding ablation study after three epochs.
Table 5.
Ablation study on TAG layer, boundary-aware attention, and boosting strategy after three training epochs.
Table 6.
Parameter sensitivity analysis for TDCN after three training epochs.