Free space exploration from a single RGB image

From a single RGB image, we can infer many properties of the scene. One of them is depth. It is a well-established problem, and we are able to produce a result reliably. In our case, we are interested in the free space in the scene, whether it is visible or invisible in the RGB image. This dissertation reports the construction of network that is able to produce a high-quality free space map from a single RGB image. This is an interesting problem that minimal current research has explored. We also investigate several strategies to achieve this goal. From our experiments, we conclude the voxel occupancy map is the most suitable because it is the most flexible and a natural extension of a pixel depth map. Utilizing 3D convolution and separating features for depth and voxel yield the best result. With several techniques that we have experimented with, the result is an improvement of about 20% on the convolutional network baseline. Producing an occupancy map by using only an RGB is an interesting topic with many potential applications. Therefore, this topic is worth investigating in the future.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Issaranon, Theerasit

Contributors dc:contributor

• Forsyth, David
• Hoiem, Derek
• Hasegawa-Johnson, Mark
• Gupta, Saurabh

• Occupancy Prediction
• Single-Image
• Space
• Voxels
• Depth

Statement dc:rights

• Copyright 2022 Theerasit Issaranon

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115735

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115735