Evaluating Segmentation Performance of Learnable Resized Models on Small Features from Sparse Point Cloud-Derived Imagery

Miguel Luis R Lagahit; Xin Liu; Haoyi Xiu; Taehoon Kim; Kyoung-Sook Kim; Masashi Matsuoka

doi:10.1109/igarss55030.2025.11242434

Home
Search
Evaluating Segmentation Performance of Learnable Resized Models on Small Features from Sparse Point Cloud-Derived Imagery

https://doi.org/10.1109/igarss55030.2025.11242434

Evaluating Segmentation Performance of Learnable Resized Models on Small Features from Sparse Point Cloud-Derived Imagery

Aug 3, 2025
Miguel Luis R Lagahit +5 more

Listen

Abstract
Literature Map
References
Similar Papers

Abstract

This study extends previous work that utilizes a learnable resizer (LR) integrated into deep learning models for segmenting sparse point cloud-derived imagery, which often suffers from poor feature representation. While prior research demonstrated the potential of LRs to enhance segmentation performance, this study evaluates their effectiveness for segmenting small and sparse features in typical scenarios where prior information is unavailable for masking. The analysis reveals that, although the LR significantly increases recall, it does not consistently or significantly improve precision. This discrepancy is attributed to the LR’s behavior, which causes target road marking features to appear as consolidated regions rather than discrete dots, leading to misclassifications in areas lacking point cloud values. Furthermore, this study also extends the ablation study regarding the impact of different resizing operations and LR placements, demonstrating that, contrary to previous work, using LR in both downsampling and upsampling phases yields better performance. Despite these challenges, the findings suggest that LRs remain valuable in retaining critical features. The study proposes that further refinement of the LR, with a focus on controlling overfitting, could improve precision while maintaining recall benefits, ultimately enhancing segmentation performance for small and sparse target features.

Similar Papers

Research Article
35
10.1007/s13243-018-0059-5

Low cost three-dimensional virtual model construction for remanufacturing industry

Oct 02, 2018
Journal of Remanufacturing
Muftooh U R Siddiqi + 7 more +7

Remanufactured products can save up to 80% of production and energy costs whilst generating lower CO2 emissions. The key success factors for remanufacturing are quality, lead-time and cost. Extensive work within the industry and the detailed analysis of the remanufacturing process has shown that component inspection has significant bearing on overall productivity. Remanufacturing lacks automation because activities are predominantly manual. Automation of remanufacturing process will not only decrease the number of non-remanufacturable components, through decreasing cost and increasing consistency in quality, but also attract industries to design for remanufacture. A digital model of the component is required to automate the disassembly process and move towards industry 4.0 and cyber physical systems. There are several expensive techniques to create a digital model, which are not feasible for the remanufacturing industry. The research paper aims to check feasibility of using Visual Structure for Motion (VFM), a relatively low cost method, to develop a 3D digital model, for automation of the automotive engine (in as received condition) disassembly process using industrial robots. These experiments assess the scientific feasibility of using Videogrammetry to acquire pre-disassembly 3D model of the engine. Multiple 2D images were acquired and processed to find matching common features. The location of the camera was calculated through the matching features, producing a three-dimensional digital representation of the captured volume. A sparse point cloud was initially created and was then converted into a dense 3D point cloud. The 3D point cloud was converted into a meshed model. 2D images were stitched together to create a virtual model of the engine with surface texture and colour. Small features were clearly visible in the 3D model.