Scene-Level Sketches for Semantic Segmentation

Citation Author(s):
Ce
Ge
Submitted by:
Ce Ge
Last updated:
Wed, 05/25/2022 - 04:22
DOI:
10.21227/bq1e-hv91
Data Format:
Research Article Link:
License:
300 Views
Categories:
0
0 ratings - Please login to submit your rating.

Abstract 

This publication contains three new datasets for scene-level sketch semantic segmentation task, namely SKY-Scene, TUB-Scene, and Freehand-Scene. SKY-Scene and TUB-Scene are synthetic datasets, where the scene layout templates were extracted from dataset SketchyScene, and the object components were adopted from dataset Sketchy. They are composed to include more sketch-specific characteristics, e.g., sparsity, abstractness, and diversity, to truly evaluate segmentation performance. Freehand-Scene are fifty real human-drawn scene sketches for practical test.

Instructions: 

SKY-Scene and TUB-Scene can be used to train, validate, and test models, whereas Freehand-Scene are collected for model testing. 

After unzipping SKY-Scene.zip or TUB-Scene.zip, there are three folders (train, val, and test) and two files (colorMap.mat and colorMap.txt). Under each folder, DRAWING_GT contains the input sketch, while CLASS_GT is the annotated semantic ground-truth. The files colorMap.mat and colorMap.txt record the class-to-color mapping; they contain the same thing but with different format.

After unzipping Freehand-Scene.zip, there is one folder test and two files colorMap.mat and colorMap.txt. Under folder test, DRAWING_GT contains the input sketch, while CLASS_GT is the annotated semantic ground-truth. The files colorMap.mat and colorMap.txt record the class-to-color mapping; they contain the same thing but with different format. This dataset is used for model evaluation.