ByLabel: A Boundary Based Semi-Automatic Image Annotation Tool - - PowerPoint PPT Presentation

ByLabel: A Boundary Based Semi-Automatic Image Annotation Tool

Xuebin Qin, Shida He, Zichen Zhang, Masood Dehghan and Martin Jagersand Department of Computing Science University of Alberta, Canada

Our Solution Introduction Results Conclusions Boundary map Region mask

We are aiming at developing a boundary based semi-automatic annotation tool to acquire pixel accurate ground truth.

Centriod Bounding box Quadrilateral Polygon

• Fig. 1 Several types of ground truth

Our Solution Introduction Results Conclusions

To acquire accurate boundary maps and region masks with light human workload, three problems have to be solved:

• P.2 Accurately

locating these control points are difficult. (Locating problem)

• P.1 Many control points are

required to describe smoothed and complex curves. (Sampling problem)

• P.3 It is hard to describe objects

with holes and objects divided by

• cclusions.

(Description problem) (a) Object with hole (b) Object divided by occlusions

• Fig. 2 Illustration of the three problems in annotation

Our Solution Introduction Results Conclusions

• Fig. 3 Workflow of our method

Our annotation tool has the following workflow:

Our Solution Introduction Results Conclusions

Feature Detection Boundaries Labeling Objects Annotation

Detect Edge Segments and Split them into Edge Fragments (Solves locating pro.) Manually select fragments to form closed boundaries (Solves sampling pro.) Group boundaries belong to the same

• bject and input the class name

(Solves description pro.)

(a) Input image (b) Edge Fragments (c) Labeled boundaries (d) Output region mask

• Fig. 4 Annotate an object step by step

Our Solution Introduction Results Conclusions

A1 A2 A3 A4 A5 B1 B2 B3 B4 B5 C1 C2 C3 C4 C5

• Fig. 5 15 Testing images in three groups

Our Solution Introduction Results Conclusions

• Table. 1. Average Clicks
• Table. 2. Average Time Costs (s)
• Table. 3. Average Error (pixel)

Our Solution Introduction Results Conclusions

(f) Instances annotation (e) Classes annotation (d) Input image (c) Bicycle (b) Earphone (a) Pedestrian

• Fig. 6 Typical annotation results

Our Solution Introduction Results Conclusions

ByLabel greatly improves the annotation efficiency and accuracy. It also reduces the annotation uncertainty and error.

Questions ??? Poster: 3C-9

ByLabel

http://webdocs.cs.ualberta.ca/~vis/bylabel/ xuebin@ualberta.ca, mj7@ualberta.ca