[Paper Review] Unmixing Convolutional Features for Crisp Edge Detection

Linxi Huan, Nan Xue+4 more|arXiv (Cornell University)|Nov 19, 2020

Advanced Neural Network Applications34 references4 citations

TL;DR

This paper proposes a context-aware tracing strategy (CATS) to improve localization accuracy in deep edge detection by unmixing convolutional features through a tracing loss and a context-aware fusion block. CATS reduces feature and side mixing in deep networks, boosting F-measure by 12% (ODS) on BSDS500 without morphological NMS when integrated with RCF using a VGG16 backbone.

ABSTRACT

This paper presents a context-aware tracing strategy (CATS) for crisp edge detection with deep edge detectors, based on an observation that the localization ambiguity of deep edge detectors is mainly caused by the mixing phenomenon of convolutional neural networks: feature mixing in edge classification and side mixing during fusing side predictions. The CATS consists of two modules: a novel tracing loss that performs feature unmixing by tracing boundaries for better side edge learning, and a context-aware fusion block that tackles the side mixing by aggregating the complementary merits of learned side edges. Experiments demonstrate that the proposed CATS can be integrated into modern deep edge detectors to improve localization accuracy. With the vanilla VGG16 backbone, in terms of BSDS500 dataset, our CATS improves the F-measure (ODS) of the RCF and BDCN deep edge detectors by 12% and 6% respectively when evaluating without using the morphological non-maximal suppression scheme for edge detection.

Motivation & Objective

• To address localization ambiguity in deep edge detectors caused by feature and side mixing in CNNs.
• To improve crispness of edge maps without relying on post-processing morphological non-maximal suppression (NMS).
• To enable high-performance edge detection using standard VGG16 backbones without architectural modifications.
• To unmix convolutional features and side predictions for better edge localization and texture suppression.
• To develop a plug-and-play module that enhances existing deep edge detectors without retraining from scratch.

Proposed method

• Introduces a novel tracing loss that performs boundary tracing to unmix features and improve side edge learning.
• Proposes a CoFusion block that adaptively fuses multi-level side edges by learning complementary attention weights.
• Uses a boundary tracing function to refine side edge predictions and reduce spatial mixing in high-level features.
• Employs a texture suppression function to enhance edge response and reduce false positives in texture regions.
• Combines the tracing loss with the CoFusion block to jointly address feature mixing and side mixing.
• Trains models end-to-end with standard initialization, integrating CATS into existing detectors like RCF, HED, and BDCN.

Experimental results

Research questions

• RQ1 Can deep edge detectors achieve sharper edge localization without morphological NMS?
• RQ2 To what extent does feature mixing in CNNs degrade edge detection performance?
• RQ3 How can side edge fusion be improved to preserve fine-grained edge details?
• RQ4 Can a lightweight, plug-and-play module enhance standard VGG16-based edge detectors?
• RQ5 Does unmixing features through boundary tracing and context-aware fusion lead to consistent performance gains?

Key findings

• CATS improves the F-measure (ODS) of RCF by 12% and BDCN by 6% on the BSDS500 dataset when evaluated without morphological NMS.
• The complete CATS achieves 0.731 ODS and 0.744 OIS on BSDS500, a 1.1% and 0.8% improvement over RCF, respectively.
• The crispness-emphasized evaluation shows a 7% improvement in ODS and 7.1% in OIS over RCF, demonstrating superior localization accuracy.
• The tracing loss alone improves ODS by 0.9% and OIS by 0.8%, while the CoFusion block alone yields minimal crispness gains.
• The CoFusion block learns to assign higher weights to lower-level side edges at boundaries and higher-level outputs in smooth regions, confirming effective side unmixing.
• Ablation studies confirm that combining both the tracing loss and CoFusion block yields the best performance, proving their complementary roles in feature and side unmixing.