Deep Structural Contour Detection

Abstract

Object contour detection is the fundamental and preprocessing step for multimedia applications such as icon generation, object segmentation, and tracking. The quality of contour prediction is of great importance in these applications since it affects the subsequent process. In this work, we aim to develop a high-performance contour detection system. We first propose a novel yet very effective loss function for contour detection. The proposed loss function is capable of penalizing the distance of contour-structure similarity between each pair of prediction and ground-truth. Moreover, to better distinguishing object contours and background textures, we introduce a novel convolutional encoder-decoder network. Within the network, we present a hyper module that captures dense connections among high-level features and produces effective semantic information. Then the information is progressively propagated and fused with low-level features. We conduct extensive experiments on the BSDS500 and Multi-Cue datasets, the results show significant improvement against the state-of-the-art competitors. We further demonstrate the benefit of our DSCD method for crowd counting.

Learning to Decode Contextual Information for Efficient Contour Detection

“…LPCB [7] explains why CNNs tend to produce blurry edges and proposes a solution to make CNNs directly predict crisp boundaries without post-processing. DSCD [6] proposes a novel loss function and three different scales of dilated convolution blocks to increase the ability of high-level feature extraction. BANet [10] presents a novel Bidirectional encoder-decoder network for accurate boundary detection.…”

“…We follow the prior works [6,7,23,36,37] to use F-score as the evaluation metric. The F-score (also known as F-measure, F1-score) can be obtained by 2 × 𝑃 × 𝑅/(𝑃 + 𝑅) where 𝑃 = 𝑇 𝑃/(𝑇 𝑃 + 𝐹 𝑃), P, TP, FP denotes the Precision, True Positive and False Positive, respectively; 𝑅 = 𝑇 𝑃/(𝑇 𝑃 + 𝐹 𝑁 ), R and FN denotes the Recall and False Negative.…”

“…Since we have already reported three ablation experiments on BSDS500, we now compare our method with top edge detectors. These SOTA methods include HED [37], RCF [22,23], CED [36], LPCB [7] and DSCD [6]. Following their training strategy, we first pretrain our method on Pascal Voc Context dataset then finetune the model on the BSDS500 training dataset.…”

“…Since it can produce important visual cues like geometry shape and semantic boundary information, it is often employed as the basic block of multimedia applications such as icon generation [34] in banners and homepages, object colorization [14], object segmentation and tracking [20]. Recently, contour detection methods equipped with deep learning techniques [6,7,23,36,37] achieved state-of-the-art performances on the datasets such as BSDS500, which demonstrates the effectiveness of deep convolutional neural networks (DCNN).…”

SketchEdit: Mask-Free Local Image Manipulation with Partial Sketches

“…2 (a). Under this inpainting framework, paired training data can be easily obtained by randomly generating masks as in general image inpainting [15,19,22,24,30,32,35,35,37,39,44,45,48,50,51,53,56,59] and extracting edges/contours in the masked regions as surrogate sketches using edge detection algorithms [1,5,6,13,21,25,26,38,43,47,49]. Then the problem can be solved by training an inpainting model to predict the original image given the masked image, mask, and sketch as input.…”