A pixel-wise annotated dataset of small overlooked indoor objects for semantic segmentation applications

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Existing datasets lack application-specific objects crucial for powered wheelchair users and indoor navigation.
• Robots require environmental understanding for navigation and object interaction.
• Deep Convolutional Neural Networks (DCNNs) often struggle with small objects.

Purpose of the Study:

• To introduce a novel dataset of annotated images for pixel classification tasks.
• To address the need for application-specific objects in datasets for powered wheelchair users and indoor robotics.
• To improve the robustness of DCNNs by including objects of various sizes and types, such as diverse door handles.

Main Methods:

• A camera mounted on a powered wheelchair recorded video footage in indoor corridors.
• Video data was annotated at the pixel level for semantic segmentation using MATLAB.
• The dataset comprises 1549 images across nine classes, including various object sizes and door handle variations.

Main Results:

• The dataset facilitates training and testing of semantic segmentation systems.
• It enables the development of more robust DCNNs capable of handling multi-size objects.
• The inclusion of diverse door handle images addresses a gap in publicly available datasets.

Conclusions:

• The dataset is valuable for powered wheelchair navigation, indoor robotics, and environmental understanding.
• It supports the creation of assistive technologies for visually impaired users.
• The dataset is publicly available to foster further research and development.