Source-linked AI summary
Bi-Directional Cascade Network for Perceptual Edge Detection
Jianzhong He, Shiliang Zhang, Ming Yang, Yanhu Shan, Tiejun Huang
TL;DR
Edge detection must handle boundaries and local details at different scales, while existing supervision and multi-scale strategies have limitations. BDCN assigns layer-specific supervision through a bi-directional cascade and uses SEM's dilated convolutions for compact multi-scale features. Across three datasets, it achieves 0.828 ODS F-measure on BSDS500, 1.3% higher than current state-of-the-art, while using a fraction of HED's parameters.
Problem
Edges at different scales are difficult to model because CNN layers receive the same supervision despite attending to different pattern scales, while deeper or pyramid-based solutions add cost.
Method
BDCN infers layer-specific edge supervision through a bi-directional cascade, while SEM uses parallel dilated convolutions to enrich multi-scale representations in a shallow network.
Results
0.828 ODS F-measure is achieved on BSDS500, 1.3% higher than current state-of-the-art, with consistent performance across BSDS500, NYUDv2, and Multicue.
Takeaways & Limitations
Learning scale-dedicated layers produces compact networks; the approach outperforms HED using only 1/6 of its parameters.
Takeaways & Limitations
The formulation notes that training without scale-differentiated supervision does not necessarily distinguish the scales represented by different layers.
Abstract
from arXiv · showhide
Exploiting multi-scale representations is critical to improve edge detection for objects at different scales. To extract edges at dramatically different scales, we propose a Bi-Directional Cascade Network (BDCN) structure, where an individual layer is supervised by labeled edges at its specific scale, rather than directly applying the same supervision to all CNN outputs. Furthermore, to enrich multi-scale representations learned by BDCN, we introduce a Scale Enhancement Module (SEM) which utilizes dilated convolution to generate multi-scale features, instead of using deeper CNNs or explicitly fusing multi-scale edge maps. These new approaches encourage the learning of multi-scale representations in different layers and detect edges that are well delineated by their scales. Learning scale dedicated layers also results in compact network with a fraction of parameters. We evaluate our method on three datasets, i.e., BSDS500, NYUDv2, and Multicue, and achieve ODS Fmeasure of 0.828, 1.3% higher than current state-of-the art on BSDS500. The code has been available at https://github.com/pkuCactus/BDCN.
1. Introduction
The paper targets edge detection across varied object scales while avoiding the parameter and training costs of deeper or explicitly multi-scale networks. It introduces BDCN for layer-specific scale supervision and SEM for compact multi-scale feature extraction, reporting strong results across three datasets.
- 1. Introduction: Edges within one image can include both object-level boundaries and meaningful local details, making multi-scale representations important for edge detection.Examples include human silhouettes and hand gestures.
- 1. Introduction: Existing approaches use deeper CNNs or image pyramids to strengthen multi-scale representations, but these strategies increase parameters, inference cost, or redundant computation.The paper identifies deeper backbones and explicit multi-level processing as costly alternatives.
- 1. Introduction: Applying one general ground-truth edge map to every CNN layer is suboptimal because different layers attend to different pattern scales.Lower layers tend toward local patterns, whereas higher layers capture object-level information with larger receptive fields.
- 1. Introduction: SEM uses parallel convolutions with different dilation rates to capture multi-scale spatial contexts without significantly increasing parameters or explicitly fusing multiple edge detections.The module is designed to enrich multi-scale representations in a shallow CNN.
- 1. Introduction: BDCN uses a bi-directional cascade to infer layer-specific supervision, training each Incremental Detection Block to focus on a proper edge scale.Adjacent higher and lower layers provide the complementary information used to infer supervision.
- 1. Introduction: 0.828 ODS F-measure is achieved on BSDS500, 1.3% higher than current state-of-the-art CED, while performance remains consistent across BSDS500, NYUDv2, and Multicue.The method also reports a better compactness–accuracy trade-off and outperforms HED using 1/6 of its parameters.
2. Related Work
The related work spans traditional, learning-based, and deep edge detection, multi-scale representation learning, and network cascade architectures. BDCN differs from prior cascades by using a bi-directional pseudo-cascade to supervise layers individually for scale-specific edge detection.
- Edge Detection: Edge detection methods include traditional operators, supervised learning with hand-crafted features, and recent deep-learning approaches.Traditional operators use changes in intensity, color, or texture, while learning-based methods use supervised models and hand-crafted features.
- Multi-Scale Representation Learning: Multi-scale representations can be constructed from image pyramids, hierarchical features, dilated convolutions, and pooling-based methods.Image pyramids may compute features independently at each scale or pass one scale's output to the next.
- Multi-Scale Representation Learning: Prior edge detectors improve multi-scale processing through Gaussian smoothing, scaled images, image pyramids, hierarchical features, or averaging predictions from multiple inputs.These approaches explore edge scale-space through several forms of feature or image rescaling.
- Network Cascade: Network cascades have been applied to classification, detection, pose estimation, and semantic segmentation using networks or pathways of differing depth or complexity.Examples include routing easy regions to shallow networks and hard regions to deeper networks, or propagating deep features to shallow layers.
- Network Cascade: Unlike previous cascades, BDCN is a bi-directional pseudo-cascade that supervises each layer individually for layer-specific edge detection.The paper presents this as an early attempt to apply cascade architecture to edge detection.
3. Proposed Methods
BDCN learns scale-specific edge detectors through complementary bidirectional supervisions, while SEM enriches each layer’s multi-scale representation without relying on a deeper network. The architecture combines these components in ID Blocks and trains side and fused predictions with class-balanced losses.
- 3.1. Formulation: Edges are decomposed into S binary maps according to the scale of their depicted objects, treating edge scale as proportional to object size.This provides the layer-specific targets needed for scale-aware edge detection.
- 3.1. Formulation: The original shared supervision gives different layers equal loss gradients, so it does not necessarily differentiate the scales represented by those layers.This makes shared supervision unsuitable for layer-specific scale learning.
- 3.1. Formulation: BDCN uses two complementary supervisions at each scale: shallow-to-deep propagation suppresses smaller-scale edges, while deep-to-shallow propagation suppresses larger-scale edges.Their two edge predictions are combined to approximate the supervision for that scale.
- 3.2. Architecture of BDCN: Each ID Block contains convolutional layers followed by SEMs, fuses their outputs, and produces P d2s and P s2d predictions for bidirectional cascade learning.The final edge map is obtained by fusing intermediate predictions with a 1×1 convolutional layer.
- 3.2. Architecture of BDCN: SEM applies parallel dilated convolutions with different dilation rates to capture multi-scale spatial contexts without reducing feature-map resolution or increasing parameters.The module uses K convolutions with rates determined by r_k = max(1, r0 × k).
- 3.3. Network Training: Different ID Blocks detect different edge scales: shallow blocks respond strongly to local details, whereas deeper blocks are more sensitive to larger-scale boundaries.Detailed zebra and butterfly body edges are detected by shallow blocks but suppressed by deeper ones.
4. Experiments
Experiments evaluate BDCN and SEM on BSDS500, NYUDv2, and Multicue, including parameter and component ablations. The method performs strongly across datasets while retaining accuracy with shallow, compact networks.
- Experimental Setup: The method is evaluated on BSDS500, NYUDv2, and Multicue using standard edge-detection protocols and metrics.BSDS500 uses augmented training data and additional PASCAL VOC Context data; NYUDv2 is evaluated with RGB, HHA, and RGB-HHA inputs; Multicue uses separate boundary and edge tasks.
- Ablation Study: K larger than 1 improves SEM performance, while overly large K or dilation rate r0 reduces performance; subsequent experiments use K=3 and r0=4.The reported explanation is that large K creates high-dimensional outputs, making edge extraction difficult.
- Ablation Study: The bi-directional cascade gives the best cascade-only performance, while SEM and the bi-directional cascade improve baseline ODS F-measure by 0.7% and 0.8%, respectively.Combining SEM with the bi-directional cascade produces the best performance among the tested component configurations.
- Comparison with Other Works: 0.828 F-measure ODS is achieved on BSDS500 with multi-scale input, while single-scale input reaches 0.820 and both outperform competing methods.The single-scale result also exceeds CED and DeepBoundary, which use multi-scale inputs.
- Comparison with Other Works: 0.748 versus 0.744 F-measure ODS is obtained against AMH-Net on NYUDv2 RGB images despite using a shallower network.The method also reports gains over RCF and HED of 0.019 and 0.028 in ODS, respectively.
- Comparison with Other Works: On Multicue, the method exceeds RCF and HED by 1.3% and 2.4% for boundary detection and 3.4% and 4.3% for edge detection.The authors also report smaller performance fluctuation than those methods.
- Compactness and Efficiency: The method remains effective with shallow networks: using 0.28M parameters and two ID Blocks, it achieves 0.766 F-measure ODS.With three ID Blocks, it outperforms HED by 0.8% using one-sixth of HED’s parameters; it runs at about 22fps with five ID Blocks.
5. Conclusions
The paper concludes that BDCN provides layer-specific, scale-focused supervision and SEM enriches multi-scale representations in a shallow network. Together, they achieve strong edge-detection performance while enabling a compact model.
- 5. Conclusions: BDCN trains each network layer with supervision tailored to a specific edge scale through a bi-directional cascade structure.SEM further enriches multi-scale representations using a shallow network.
- 5. Conclusions: 1/6 of HED’s parameters are sufficient for the approach to outperform HED, demonstrating a favorable compactness-accuracy trade-off.The conclusion attributes this compactness to learning scale-dedicated layers.