Skip to main content
QUICK REVIEW

[논문 리뷰] Bi-Directional Cascade Network for Perceptual Edge Detection

Jianzhong He, Shiliang Zhang|arXiv (Cornell University)|2019. 02. 28.
Advanced Neural Network ApplicationsComputer Science참고 문헌 61인용 수 38
한 줄 요약

BDCN은 계층별 감독이 있는 양방향 cascade와 Scale Enhancement Module을 도입해 얕은 네트워크에서 다중 스케일 에지 표현을 학습하고, BSDS500, NYUDv2, Multicue에서 에지 검출의 최첨단 성능을 달성합니다.

ABSTRACT

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.

연구 동기 및 목표

  • Motivate multi-scale edge detection to handle edges at varying object scales.
  • Propose a Bi-Directional Cascade Network (BDCN) to supervise each layer for scale-specific edges.
  • Introduce a Scale Enhancement Module (SEM) using dilated convolutions to enrich multi-scale representations.
  • Achieve competitive edge detection performance with a shallow network.
  • Show improvements across BSDS500, NYUDv2, and Multicue datasets.

제안 방법

  • Introduce Incremental Detection Block (ID Block) with embedded SEMs to produce scale-targeted edge predictions.
  • Use a bi-directional cascade to propagate edge predictions from shallow to deep and from deep to shallow layers during training (S2D and D2S); combine via a fusion layer.
  • Train each ID Block with two layer-specific supervisions (s2d and d2s) and an overall fusion loss.
  • SEM applies multiple parallel dilated convolutions with varied dilation rates to capture multi-scale context within a block.
  • Formulate a multi-term loss: L = w_side * L_side + w_fuse * L_fuse, with side losses L_side comprising P_s^{d2s} vs Y_s^{d2s} and P_s^{s2d} vs Y_s^{s2d}.
  • Ground truth is decomposed into scale-specific edge maps and complementary supervisions Y_s^{s2d} and Y_s^{d2s} to encourage scale-dedicated learning.

실험 결과

연구 질문

  • RQ1Can layer-specific, scale-dedicated supervision improve edge detection across multiple scales?
  • RQ2Does a bi-directional cascade enable effective scale specialization without heavy reliance on deeper networks?
  • RQ3Can a shallow network augmented with SEM achieve competitive or superior edge detection performance?
  • RQ4How do SEM parameters (dilation rates and number of dilations) affect multi-scale edge learning?
  • RQ5Do BNDCN variants (with/without SEM, with various cascade directions) differ in performance on standard datasets?

주요 결과

  • BDCN with SEM achieves state-of-the-art ODS F-measure on BSDS500 (0.828) surpassing prior methods by 1.3%.
  • BDCN outperforms human perception on BSDS500 with trainval data (ODS 0.828 vs 0.803).
  • BDCN attains competitive results on NYUDv2 and Multicue, often outperforming competing methods in ODS, OIS, and AP metrics.
  • With fewer parameters, BDCLN (BDCN) achieves strong performance, e.g., 5-block model reaches about 16M parameters and 0.82 ODS on BSDS500.
  • Ablation shows that combining SEM with bi-directional cascade yields the best results, and SEM is particularly critical for multi-scale representation in shallow networks.
  • BDCN shows robustness to depth reduction, maintaining competitive ODS with a fraction of parameters compared to HED and RCF.

이 논문 읽기

누빈트 뷰어에서 원문을 열고, 읽으면서 궁금한 곳을 바로 물어보세요.

누빈트에서 이 논문 읽기

이 리뷰는 AI가 논문 원문을 바탕으로 만들었습니다. 인용하기 전에 원문을 확인해 주세요.