Understanding Dice Loss for Crisp Boundary Detection

A Far Better Alternative to Cross Entropy Loss for Boundary Detection Tasks in Computer Vision

Shuchen Du
AI Salon

--

Photo by Brittney Dowell on Unsplash

In deep learning and computer vision, people are working hard on feature extraction to output meaningful representations for various kinds of vision tasks. In some tasks, we only focus on geometry of the objects, regardless of colors, textures and illuminations etc. That ‘s where boundary detection comes in.

Problem Setting

Fig. 1: a boundary detection example [Deng. et al.]

As the name implies, boundary detection is the task that detect object boundaries from images. It is an ill-posed problem as ambiguity exists in the problem setting itself. As shown in Fig. 1, for an indoor room image (left), the ground truth (middle) defines ground truth object boundaries inside the room, and a prediction (right) estimates object boundaries of the room. However, we can see that estimated boundaries are far more than ground truth, including unnecessary boundary lines from room layouts, curtains, and even sofa textures. Extracting clean and meaningful object boundaries is not easy.

Naive Methods

--

--