Devil Is in the Edges: Learning Semantic Boundaries From Noisy Annotations

Abstract

We tackle the problem of semantic boundary prediction, which aims to identify pixels that belong to object(class) boundaries. We notice that relevant datasets consist of a significant level of label noise, reflecting the fact that precise annotations are laborious to get and thus annotators trade-off quality with efficiency. We aim to learn sharp and precise semantic boundaries by explicitly reasoning about annotation noise during training. We propose a simple new layer and loss that can be used with existing learning-based boundary detectors. Our layer/loss enforces the detector to predict a maximum response along the normal direction at an edge, while also regularizing its direction. We further reason about true object boundaries during training using a level set formulation, which allows the network to learn from misaligned labels in an end-to-end fashion. Experiments show that we improve over the CASENet [36] backbone network by more than 4% in terms of MF(ODS) and 18.61% in terms of AP, outperforming all current state-ofthe-art methods including those that deal with alignment. Furthermore, we show that our learned network can be used to significantly improve coarse segmentation labels, lending itself as an efficient way to label new data.

Neural Turtle Graphics for Modeling City Road Layouts

“…• Stronger CNNs: We explore more powerful CNN architectures. We train an FCN with a ResNet backbone [20,28], as well as a CNN using DLA [42] with STEAL [3]. To obtain the graph we use standard thinning and geodesic sorting.…”

On Improving the Training of Models for the Semantic Segmentation of Benthic Communities from Orthographic Imagery

“…Following a similar concept, in [34], the authors subdivide the background class into two separate sets with the instance boundaries proximity and weight them differently. Related work [35] proposes a custom combination of layer and loss to learn sharper and more adherent semantic boundaries. This flexible solution can be easily adapted to work with different semantic segmentation architecture.…”

Detecting the occluding contours of the uterus to automatise augmented laparoscopy: score, loss, dataset, evaluation and user study

“…Evaluation score. The principle of the score from [16,9] is widely used in semantic edge detection [26,2,25]. The score is based on precision-recall obtained by matching responses and true edges.…”

“…For both, the best results are currently obtained with CNNs. In semantic edge detection, the task is to detect the boundary of multiple specific objects [26,2,25,6,13]. Weighted cross-entropy is commonly used to compensate the imbalanced distribution between the edge and non-edge classes over the image.…”

“…The second missing part for OC2D is the detector itself, specifically the loss to train a CNN end-to-end. Using a CNN is a natural approach, as in related tasks [26,2,25,6,17,22,23]. These methods do not address OC2D specifically but reveal the important potential problem of thick responses [2,25,6].…”