Source-linked AI summary
Dense Extreme Inception Network for Edge Detection
Xavier Soria, Angel Sappa, Patricio Humanante, Arash Akbarinia
TL;DR
The paper addresses the mismatch between commonly used boundary-oriented datasets and the requirements of edge detection, where fine-scale edges matter. It introduces the BIPED dataset and DexiNed, an end-to-end network trained from scratch, and reports strong generalization across datasets. The authors conclude that deep edge detectors can be trained from scratch and generalized to new scenes.
Problem
Current edge-detection models are shaped by datasets whose annotations often represent boundaries rather than all true low-level edges, limiting robust generalization to new scenes.
Method
The paper develops the BIPED edge dataset and DexiNed, an end-to-end architecture with dense skip-connections that predicts edge maps without pre-trained weights.
Results
DexiNed trained on BIPEDv2 generalizes to fine-scale edges in BSDS images and achieves higher performance than other algorithms on the presented edge dataset.
Takeaways & Limitations
The findings support training deep edge-detection models from scratch and evaluating them across datasets, while suggesting exploration of smaller networks.
Takeaways & Limitations
Evaluation on NYUD is affected by poor annotations in some images, where missing edges reduce the quality of deep-learning edge detection.
Abstract
from arXiv · showhide
<<<This is a pre-acceptance version, please, go through Pattern Recognition Journal on Sciencedirect to read the final version>>>. Edge detection is the basis of many computer vision applications. State of the art predominantly relies on deep learning with two decisive factors: dataset content and network's architecture. Most of the publicly available datasets are not curated for edge detection tasks. Here, we offer a solution to this constraint. First, we argue that edges, contours and boundaries, despite their overlaps, are three distinct visual features requiring separate benchmark datasets. To this end, we present a new dataset of edges. Second, we propose a novel architecture, termed Dense Extreme Inception Network for Edge Detection (DexiNed), that can be trained from scratch without any pre-trained weights. DexiNed outperforms other algorithms in the presented dataset. It also generalizes well to other datasets without any fine-tuning. The higher quality of DexiNed is also perceptually evident thanks to the sharper and finer edges it outputs.
1. Introduction
The paper identifies training annotations as a central limitation for edge detection and distinguishes edges from contours and boundaries. It introduces the BIPED dataset and DexiNed to improve fine-scale edge detection and cross-dataset generalization.
- Dataset motivation: BSDS annotations mainly represent high-level object boundaries, making them unsuitable as an independent benchmark for low-level edge detection.The paper argues that separate datasets are needed for edge, contour, and boundary detection.
- Proposed solution: DexiNed trained on BIPEDv2 generalizes to fine-scale edges in BSDS images, unlike the same network trained on BSDS.The paper attributes this robustness to careful annotation of edges within and across objects in BIPEDv2.
- Dataset motivation: Consensus labeling in BSDS can remove true edges, including zebra stripes and details in a red car.Pixels marked by more than two participants are retained, while other annotations are discarded.
- Contributions: BIPED is a benchmark dataset designed specifically for edge detection, with increased annotation coverage for fine-scale edges.The dataset is intended to support accurate training and evaluation of edge detection algorithms.
- Contributions: DexiNed is an end-to-end architecture with parallel skip-connections that learns edges without ImageNet pre-trained weights.The paper also describes architecture analysis and modified BDCN-based loss functions.
- Contributions: The study establishes a benchmark by interchanging training and validation sets between BIPED and MDBD across four state-of-the-art networks.This design evaluates generalization across two datasets of edge detection.
2. Related Work
Prior edge-detection research spans low-level operators, biologically inspired models, classical learning, and deep learning. Modern CNN methods generally outperform traditional approaches but often rely on VGG16 pretraining and may produce coarse edges.
- Taxonomy: Edge-detection algorithms are grouped into low-level feature methods, brain-inspired methods, classical learning-based methods, and deep learning methods.The paper presents these as four broad categories in its related-work review.
- Low-level methods: Early low-level methods use first- or second-order derivatives, smoothing, thresholding, or zero-crossing extraction to detect edges.Canny combines three key processes in a later classical formulation.
- Brain-inspired methods: Brain-inspired approaches model mechanisms of biological vision, including retinal processing, receptive fields, spatial facilitation, and surround inhibition.Several methods use Gaussian derivatives or Gabor filters to represent these mechanisms.
- Classical learning-based methods: Classical learning-based methods extract brightness, color, and texture gradients before training classifiers for probabilistic boundary detection.The paper identifies Pb as an early learning-based approach.
- Deep learning methods: Deep learning methods commonly use VGG16 convolutional layers pretrained on ImageNet, then fine-tune them on boundary-detection datasets.HED uses multi-scale outputs and deep supervision, while related models extend this design.
- Deep learning methods: Although deep models outperform traditional methods on standard datasets, their predicted edge maps can remain coarse rather than sharp.GAN-based methods and CATS are discussed as approaches for improving edge-map crispness.
3. Proposed Approach
DexiNed is an end-to-end architecture that combines a Dense Extreme Inception Network with an upsampling network to preserve and fuse edge information across scales. It is designed for training from scratch and uses a weighted cross-entropy loss for predicted edge maps.
- Architecture: DexiNed enables end-to-end training without pretrained object-detection weights by using parallel skip-connections to retain edge features across layers.The design was motivated by the loss of shallow edge features in deeper layers.
- Architecture: The architecture contains six encoder-like blocks whose feature maps are separately upsampled into intermediate edge maps, then concatenated and fused into one output.The flowchart identifies the Dense Extreme Inception Network and USNet as the two main building blocks.
- Architecture: Parallel skip-connections link blocks and sub-blocks, while additional connections average and sum intermediate outputs before later processing.The second skip-connections operate from Block-3 onward alongside direct max-pooling paths.
- Upsampling: USNet uses conditional convolutional and deconvolutional blocks to upsample feature maps until they reach the required ground-truth scale.One block is iterated for scaling, after which the other processes the upsampled features.
- Upsampling: Three upsampling strategies—bilinear interpolation, sub-pixel convolution, and transpose convolution—are considered because upsampling influences thin-edge generation.The paper evaluates these alternatives as part of DexiNed’s design.
- Loss function: The model trains with a weighted cross-entropy objective over predicted edge maps, using sigmoid outputs and hyperparameters to balance positive and negative samples.The network produces a set of predicted edge maps, and the loss weight and class-balancing parameters are defined in the training formulation.
4. Datasets
The paper introduces and updates BIPED as an edge-focused dataset, then evaluates DexiNed across datasets intended for edge, contour, and boundary detection. The benchmark collection varies substantially in annotation purpose, image content, modality, and scale.
- BIPED: BIPED contains 250 urban real-world images at 1280 × 720 resolution with ground-truth edge maps generated using Labelbox.The dataset is an updated version of Barcelona Images for Perceptual Edge Detection.
- BIPED: BIPED annotations are produced through expert labeling, administrative review, HED-based checking, correction, and additional annotation of fine-scale details.The same administrator applies consistent criteria throughout the process.
- Evaluation datasets: DexiNed is evaluated on MDBD, BSDS500, BSDS300, NYUD, PASCAL-CONTEXT, CID, DCD, and BIPED to compare performance across edge, contour, and boundary datasets.MDBD is identified as most relevant because its annotations correspond to true edges.
- Evaluation datasets: CID contains 40 grayscale 512 × 512 images with contour ground truths, making it difficult for deep-learning methods because of its small size and missed annotations.DexiNed is evaluated on the entire dataset, consistent with prior work.
- Evaluation datasets: BSDS300 has 300 images split into 200 training and 100 validation images, while BSDS500 adds 200 testing images and is primarily intended for segmentation and boundary detection.The paper notes that some BSDS images are not well annotated for edge detection.
- Evaluation datasets: NYUD provides 1449 RGBD images from 464 indoor scenarios, while PASCAL contains 11530 annotated images but only 505 are used to test DexiNed.The paper uses only NYUD’s testing set and a randomly selected subset of PASCAL images.
5. Experimental Results
Experiments evaluate DexiNed’s metrics, training setup, architectural choices, cross-dataset performance, and qualitative edge maps. The results support strong same-dataset performance, improved generalization from BIPED, and cleaner predictions without pre-training.
- Evaluation Protocol: The study evaluates edge detectors directly against human-drawn edges using ODS, OIS, and AP.ODS uses a global dataset threshold, OIS uses an image-specific threshold, and AP measures average precision.
- Implementation: DexiNed is trained from scratch without pre-trained weights and converges after 9 epochs using Adam with a 10^-4 learning rate.The learning rate decreases at 10 and 15 epochs, and the reported weight decay is 10^-8.
- Architecture Settings: Transpose convolution with trainable kernels is selected for upsampling, while the architecture merges outputs from multiple DexiNed stages.The selected upsampling method was reported to perform best in preliminary comparisons.
- Ablation Study: Two skip-connections improve DexiNed’s performance over configurations using zero or one skip-connection.The configurations are denoted DexiNed0C, DexiNed1C, and DexiNed2C according to their number of skip-connections.
- Ablation Study: The modified BDCNloss2 function outperforms the other tested losses by almost 1%, with best performance reached at 11 training epochs.Subsequent comparisons use DexiNed2C, 11 epochs, and BDCNloss2.
- Quantitative Comparison: DexiNed achieves the best ODS, OIS, and AP when trained and tested on the same dataset, and also performs best when trained on BIPED and evaluated on MDBD.Across approaches, performance loss averages 3,74% from BIPED to MDBD versus 7,14% from MDBD to BIPED.
- Qualitative Comparison: Qualitatively, DexiNed predicts more edges than CATS while remaining cleaner than its counterparts and detecting fine structures across other datasets.On PASCAL, DexiNed detects several floor edges that the other compared models do not detect.
6. Conclusion
The paper contributes an edge-specific benchmark dataset and a parallel-skip-connection network that generalizes across datasets without ImageNet pre-training. Its results demonstrate end-to-end training from scratch and motivate exploration of smaller edge-detection networks.
- 6. Conclusion: The paper proposes a benchmark dataset designed specifically for edge detection and a robust model for generalization to new scenes.The authors frame these as the paper’s two primary contributions.
- 6. Conclusion: DexiNed uses parallel skip-connections to learn edges without ImageNet pre-trained weights.The architecture is presented as an end-to-end deep learning approach.
- 6. Conclusion: Training on one dataset and evaluating on other benchmark datasets demonstrates the approach’s generalization power.The conclusion states that the model can be trained from scratch in an end-to-end fashion.
- 6. Conclusion: The findings open an opportunity to explore smaller edge-detection networks by reducing the number of hyperparameter settings.This is presented as a future direction enabled by the reported findings.