Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks
Dissertation (MEng (Computer Engineering))--University of Pretoria, 2025.
Saved in:
| Other Authors: | |
|---|---|
| Format: | Thesis |
| Language: | en_US |
| Published: |
University of Pretoria
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1869484010467491840 |
|---|---|
| access_status_str | Open Access |
| author2 | Grobler, Hans |
| author_browse | Grobler, Hans |
| author_facet | Grobler, Hans |
| collection | Thesis |
| dc_rights_str_mv | © 2024 University of Pretoria. All rights reserved. The copyright in this work vests in the University of Pretoria. No part of this work may be reproduced or transmitted in any form or by any means, without the prior written permission of the University of Pretoria. |
| description | Dissertation (MEng (Computer Engineering))--University of Pretoria, 2025. |
| format | Thesis |
| id | oai:repository.up.ac.za:2263/107858 |
| institution | University of Pretoria (South Africa) |
| language | en_US |
| last_indexed | 2026-07-01T04:08:05.715Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from UPSpace — University of Pretoria Institutional Repository |
| publishDate | 2026 |
| publishDateRange | 2026 |
| publishDateSort | 2026 |
| publisher | University of Pretoria |
| publisherStr | University of Pretoria |
| record_format | dspace |
| source_str | UPSpace — University of Pretoria Institutional Repository |
| spelling | oai:repository.up.ac.za:2263/107858 Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks Grobler, Hans u19012072@tuks.co.za Van Eeden, Janco UCTD Sustainable Development Goals (SDGs) 3D object detection 3D reconstruction Convolutional neural network Radiance field Volume rendering Dissertation (MEng (Computer Engineering))--University of Pretoria, 2025. Indoor 3D object detection is a crucial computer vision task that enhances scene understanding and interpretation. Improving the efficiency and reliability of perception systems is essential for applications in robot perception, augmented reality, and virtual reality. Although 2D object detection methods have made significant progress, 3D environments present unique challenges that require specialized approaches. Despite the advancements in the 3D-domain research, existing methods rely on accurate depth information, which is often obtained through supplementary depth sensors or depth-estimation algorithms. RGB-only 3D object detection is inherently ill-posed due to ambiguous depth cues, occlusion, illumination, and camera motion. Advancements in radiance field representations, such as neural radiance fields (NeRFs) and 3D Gaussian splatting (3DGS), have shown remarkable success in novel view synthesis (NVS); their potential for 3D object detection remains underexplored. Recent methods rely on custom radiance field optimization pipelines that require significant memory overhead and compromise the modularity and generalizability of the approach. In this research, a CNN-based 3D object detection pipeline is developed that operates directly on discrete radiance field representations and thereby aims to address the generalizability and efficiency limitations of existing methods. The proposed two-stage detection system first reconstructs scene geometry using implicit or explicit radiance field representations and then performs class-agnostic 3D object detection in indoor environments. NVS approaches based on multilayer perceptrons (MLPs), 3D Gaussians, and sparse voxel-grids are extensively evaluated across multiple criteria to determine the most effective representation for 3D object detection. Two variants of the detector are developed - dense variants for superior feature extraction and sparse variants optimizing the balance between computational efficiency and detection performance. The embedded RGB-D features are extracted from the radiance field representations via efficient dense and sparse voxel-grids. The non-maximum suppression (NMS) algorithm is optimized iteratively to significantly reduce inference speed. The proposed method is validated on three challenging indoor multi-view RGB datasets (ScanNet, Hypersim, and ARKitScenes) and evaluated against state-of-the-art (SOTA) RGB-only, NeRF-based, and point-based object detection approaches. Using a smaller subset of 90 scenes from the multi-view RGB ScanNet dataset, the designed detector achieves a recall (Rec) and average precision (AP) scores of 90.0 and 45.5, respectively, at a 25% intersection-over-union (IoU) threshold. The detector’s AP performance surpasses that of existing NeRF-based detectors by approximately 17%-26% at the same IoU overlap, while performing 13% worse than existing RGB-D point-based detectors. The sparse variants, based on 3DGS and sparse voxels, provide the optimal balance between computational efficiency and detection accuracy, achieving an average inference speed of 3.4 seconds per scene on a low-end graphics processing unit (GPU) with 6GB of memory. Cross-dataset generalization experiments revealed robustness across different capture scenarios and RGB camera types, indicating that the approach captures fundamental geometric relationships. The study also established the structural similarity index (SSIM) as a reliable predictor of 3D detection performance, revealing a strong correlation between radiance field quality and detection accuracy. The modular feature extraction framework facilitates seamless adaptation to emerging radiance field representations, while the two-stage design preserves the photorealism of NVS. Electrical, Electronic and Computer Engineering MEng (Computer Engineering) Unrestricted Faculty of Engineering, Built Environment and Information Technology None 2026-02-05T07:55:47Z 2026-02-05T07:55:47Z 2026-04 2025-12 Dissertation * April 2026 http://hdl.handle.net/2263/107858 en_US © 2024 University of Pretoria. All rights reserved. The copyright in this work vests in the University of Pretoria. No part of this work may be reproduced or transmitted in any form or by any means, without the prior written permission of the University of Pretoria. application/pdf University of Pretoria |
| spellingShingle | UCTD Sustainable Development Goals (SDGs) 3D object detection 3D reconstruction Convolutional neural network Radiance field Volume rendering Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title | Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title_full | Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title_fullStr | Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title_full_unstemmed | Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title_short | Discrete radiance field representations for 3D object detection via sparse-dense convolutional neural networks |
| title_sort | discrete radiance field representations for 3d object detection via sparse dense convolutional neural networks |
| topic | UCTD Sustainable Development Goals (SDGs) 3D object detection 3D reconstruction Convolutional neural network Radiance field Volume rendering |
| url | http://hdl.handle.net/2263/107858 |