Object detection of urban trees for scene generation

Scene generation is the recreation of an environment. The detection and localization of trees is an essential part of automating 3D visualization tools which generate scenes. Scene generation is useful in many scenarios and experiments where the outcome of the experiment is affected by the environme...

Full description

Saved in:
Bibliographic Details
Main Author: Koh, Mitchell Yiang Dhee
Other Authors: Lee Bu Sung, Francis
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/163051
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Scene generation is the recreation of an environment. The detection and localization of trees is an essential part of automating 3D visualization tools which generate scenes. Scene generation is useful in many scenarios and experiments where the outcome of the experiment is affected by the environment. Some studies which require scene generation include studying the rate of deforestation and studying the effects of temperature in an urban landscape. For this project, the aim of tree detection is to be able to help recreate an urban environment to study noise pollution and how noise is reduced by the presence of trees. This final year project reports on the detection of trees from street view images, with the use of deep learning models. In the exploratory phase, different computer vision tasks were compared against each other to determine the most appropriate task for the project. In the preparation phase, to generate input for the image detection model, the preparation of a dataset was conducted. This process included sourcing for suitable datasets, exploration of different annotation tools, and deciding the type of image recognition task to be used. Finally for the experiment phase. The experiment task chosen was instance segmentation where the objective of the model was to detect the masks of each individual tree. Several proposals were made to improve the model performance. The wide variety of tree species, clarity and overlapping of trees and limited dataset size were some challenges faced when training the model to detect each instance of a tree from street view images.