cs.CVMar 21, 2026
SaveMEMO: Human-like Crisp Edge Detection Using Masked Edge Prediction
Organizations: Department of Computer and Information Science, University of Macau, Macau, China · Capital One, USA
Abstract
Learning-based edge detection models trained with cross-entropy loss often suffer from thick edge predictions, which deviate from the crisp, single-pixel annotations typically provided by humans. While previous approaches to achieving crisp edges have focused on designing specialized loss functions or modifying network architectures, we show that a carefully designed training and inference strategy alone is sufficient to achieve human-like edge quality. In this work, we introduce the Masked Edge Prediction MOdel (MEMO), which produces both accurate and crisp edges using only cross-entropy loss. We first construct a large-scale synthetic edge dataset to pre-train MEMO, enhancing its generalization ability. Subsequent fine-tuning on downstream datasets requires only a lightweight module comprising 1.2% additional parameters. During training, MEMO learns to predict edges under varying ratios of input masking. A key insight guiding our inference is that thick edge predictions typically exhibit a confidence gradient: high in the center and lower toward the boundaries. Leveraging this, we propose a novel progressive prediction strategy that sequentially finalizes edge predictions in order of prediction confidence, resulting in thinner and more precise contours. Our method achieves visually appealing, post-processing-free, human-like edge maps and outperforms prior methods on crispness-aware evaluations.
Figures & tables
Figure 1: Overview of MEMO’s architecture. MEMO is trained in two stages. In the first stage, we train the masked edge encoder and edge decoder on a large-scale synthetic edge dataset while keeping the pre-trained image encoder frozen. In the second stage, we fine-tune on downstream datasets by injecting LoRA adapters into the edge encoder and decoder, while keeping all other components fixed. During inference, MEMO begins with a fully masked edge map and iteratively completes it based on a confidence-ordered strategy. At each step, the high confident predictions are finalized, and the remaining pixels are remasked for the next iteration according to the LocMax strategy in Section 2.3 . For better visualization, block-wise masks are illustrated, although masking is applied pixel-wise in practice.
Figure 2: With very few steps, MEMO produces ambiguous edge maps. As the number of steps increases, the edges become sharper and more structurally coherent. Notably, using just 10 steps already yields visually crisp results.
Figure 3: Multi-granularity prediction results of MEMO under varying granularity scales . As the scale increases from to , MEMO predicts increasingly dense edge maps with finer structural details.
Figure 4: Qualitative comparison between MEMO and baseline methods on the BSDS dataset. MEMO produces crisp and structurally accurate edges, closely resembling human-style annotations. In contrast, baseline methods often produce thick or ambiguous edges, which increases the risk of merging closely spaced edges into a single contour. This leads to a loss of fine details and reduced separation between adjacent edges. MEMO avoids this by maintaining sharp and well-localized predictions, even in regions with densely packed edges.
| Methods | SEval | CEval | AC | ||
| ODS | OIS | ODS | OIS | ||
| Canny [ 2 ] | 0.610 | 0.650 | - | - | - |
| SE [ 4 ] | 0.743 | 0.764 | - | - | - |
| OEF [ 7 ] | 0.746 | 0.770 | - | - | - |
| DeepCntr [ 24 ] | 0.757 | 0.776 | - | - | - |
| DeepBdry [ 16 ] | 0.789 | 0.811 | - | - | - |
Table 1: Standard and Crispness-aware evaluation on BSDS. All results are reported using a single-scale prediction. C*: Best CEval when . AC*: Best AC when .
| Methods | SEval | CEval | AC | ||
| ODS | OIS | ODS | OIS | ||
| HED [ 30 ] | 0.829 | 0.847 | - | - | - |
| RCF [ 17 ] | 0.843 | 0.859 | 0.640 | 0.647 | 0.212 |
| BDCN [ 8 ] | 0.839 | 0.854 | 0.629 | 0.637 | 0.238 |
| DexiNed [ 25 ] | 0.859 | 0.867 | 0.764 | 0.776 | 0.295 |
| DiffEdge [ 31 ] | 0.899 | 0.901 | 0.856 | 0.863 | 0.849 |
Table 2: Standard and Crispness-aware evaluation on BIPED. C*: Best CEval when . AC*: Best AC when .
| Methods | SEval | CEval | AC | ||
| ODS | OIS | ODS | OIS | ||
| HED [ 30 ] | 0.851 | 0.864 | - | - | - |
| RCF [ 17 ] | 0.851 | 0.862 | 0.674 | 0.691 | 0.203 |
| BDCN [ 8 ] | 0.891 | 0.898 | 0.676 | 0.679 | 0.249 |
| DexiNed [ 25 ] | 0.872 | 0.881 | 0.764 | 0.776 | 0.274 |
| EDTER [ 22 ] | 0.894 | 0.900 | 0.735 | 0.747 | 0.196 |
Table 3: Standard and Crispness-aware evaluation on Multicue. C*: Best CEval when . AC*: Best AC when .
Figure 5: In regions containing dense and fine‑scale structures, such as text strokes and tree boundaries, DiffEdge [ 31 ] exhibits inconsistent localization and produces fragmented or overly softened contours. In contrast, MEMO preserves high‑frequency edge details and yields coherent, well‑separated contours that more closely align with the ground‑truth annotations.
Table 4: Multi-Granularity Evaluation on BSDS
| Methods | AC | FID | LPIPS |
| MuGE [ 34 ] | 0.296 | 115.89 | 0.456 |
| DiffEdge [ 31 ] | 0.476 | 89.96 | 0.300 |
| SAUGE [ 18 ] | 0.197 | 243.01 | 0.651 |
| CATS [ 14 ] | 0.333 | 135.89 | 0.421 |
| Refined Lbl [ 32 ] | 0.424 | 135.29 | 0.370 |
| MEMO (C*) | 0.663 | 83.95 | 0.282 |
Table 5: Quantitative comparison of visual similarity to human annotations. MEMO outperforms both high-performance edge detectors [ 34 , 31 , 18 ] and crispness-focused methods [ 14 , 32 ] .
| Inf. Steps | 5 | 10 | 20 | 40 | Full | |
| SEval | ODS | 0.855 | 0.854 | 0.847 | 0.846 | 0.846 |
| OIS | 0.861 | 0.861 | 0.854 | 0.852 | 0.852 | |
| CEval | ODS | 0.835 | 0.836 | 0.838 | 0.841 | 0.842 |
| OIS | 0.840 | 0.841 | 0.842 | 0.843 | 0.843 | |
| AC | 0.594 | 0.663 | 0.737 | 0.828 | 0.840 | |
| Inf. Time | (sec.) | 0.62 | 1.33 | 2.77 | 5.63 | 10.46 1.03 |
Table 6: MEMO’s performance when using different inference unmasking steps. Inference time is measured on Nvidia 3090 GPU.
| Inference Strategy | Random | TopK | LocMax | |
| SEval. | ODS | 0.819 | 0.825 | 0.854 |
| OIS | 0.829 | 0.843 | 0.861 | |
| CEval. | ODS | 0.794 | 0.715 | 0.836 |
| OIS | 0.804 | 0.720 | 0.841 | |
| AC | 0.671 | 0.510 | 0.663 | |
Table 7: MEMO’s performance using different inference strategies
Figure 6: Qualitative comparison of MEMO when using different unmasking strategy. The random unmasking produces low quality prediction with discontinued edges. The TopK unmasking can still cause concentrated edges. The proposed LocMax unmasking scheme yields crisp and good coherent edge predictions.
Figure 7: When trained only on real data, MEMO can produce duplicated or echoed contours in regions with ambiguous structure. Pre-training on synthetic data suppresses these multi-line artifacts and yields clean, single-edge predictions.
| Training Set | Synth. | Real | Both | |
| SEval. | ODS | 0.729 | 0.804 | 0.854 |
| OIS | 0.748 | 0.810 | 0.861 | |
| CEval. | ODS | 0.725 | 0.799 | 0.836 |
| OIS | 0.726 | 0.803 | 0.841 | |
| AC | 0.765 | 0.619 | 0.663 | |
Table 8: Performance of MEMO when trained on synthetic-only, real-only, and both dataset.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Edge predictions under varying granularity scale values. As the scale increases, more edges are progressively revealed. MEMO maintains high edge quality across a wide range of scales, but excessively large values may introduce spurious edges. This demonstrates MEMO’s flexible granularity control beyond pre-defined ranges used in prior work.
Figure 9: Distribution of edge pixel finalization steps across the inference process. Left: Histogram showing the percentage of edge pixels finalized at different unmasking step ranges. Right: Cumulative proportion of finalized pixels by step count.
Figure 10: Intermediate predictions at different inference stages of MEMO. The second column shows edge predictions after the first inference pass, where all pixels are visualized. These predictions, similar to conventional methods, reveal thick and ambiguous prediction. However, the edges of different ambiguity converge at different speed. The last two columns show the finalized edge pixels after 20 steps and full inference, respectively. Less ambiguous boundaries usually converge early (green boxes), while highly uncertain regions (red boxes) require more iterations to stabilize.
Figure 11: With fewer steps (5 steps), ambiguous edges are partially predicted with low confidence, resulting in faint contours. With full-step inference, MEMO finalizes these edges, leading to sharper but potentially unmatched predictions due to missing ground-truth labels. Although evaluation protocol may penalize these as false positives, they remain visually plausible from a human perspective.
Figure 12: Qualitative comparison between MEMO without post-processing and baseline methods with post-processing. Although post-processing enhances benchmark scores, it does not guarantee perceptually faithful edge maps. Baseline predictions often exhibit duplicated, broken, or jittery edges, particularly in fine structures like the antenna and rooflines, due to the inherent ambiguity in thick or overlapping predictions. In contrast, MEMO directly predicts crisp and coherent edge structures, accurately capturing both global layouts and fine details without the need for post-processing.
Figure 13: Example of synthetic edge dataset. Images are cropped to for visualization.
Figure 14: Example of synthetic edge dataset. Images are cropped to for visualization.