Focal Combo Loss for Improved Road Marking Extraction of Sparse Mobile LiDAR Scanning Point Cloud-Derived Images Using Convolutional Neural Networks

Lagahit, Miguel Luis R.; Matsuoka, M.

doi:10.3390/rs15030597

Cited by 6 publications

(8 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As mentioned earlier, the loss function is one of the components of the model training procedure that takes the differences between initial predictions and labeled ground truth to adjust the weights inside the network. In this paper, we evaluate Focal Combo (FC) loss, as shown in Equation 1 (Lagahit et al, 2023).…”

Section: Loss Functionsmentioning

confidence: 99%

Investigating Loss Functions for Segmenting and Detecting Ships on Sar Imagery

Lagahit,

Matsuoka

2024

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.

View full text Add to dashboard Cite

Abstract. In accordance with the United Nations (UN) Sustainable Development Goal (SDG) 16: Peace, Justice, and Strong Institutions, this study explores ship monitoring through the use of Synthetic Aperture Radar (SAR) for its potential applications to economic and security purposes. One method to extract ships through SAR-derived imagery is to employ the use of convolutional neural networks (CNN). However, the extraction of small features continues to be a challenging task for CNNs. To improve the performance in such cases, one way is to employ the use of an appropriate loss function, which helps guide the CNN model during training. In this paper, Focal Combo (FC) loss, a recent loss function designed for extreme class imbalance, will be investigated to analyze its effects when applied to ship extraction. In doing so, this paper also presents a thorough comparison of existing loss functions in their capability to segment and detect ships on SAR imagery. Making use of the U-Net model, our results demonstrate that by using FC loss we can observe an increase in segmentation of about 9% in terms of f3-score and a decrease in missed detections by about 17 ships (after post-processing) when compared to cross-entropy loss. Unfortunately, it has also shown a significant drop in precision of about 35% resulting in an additional 270 ships being incorrectly detected in the background. In future work, varying CNN models shall be tested to see if the pattern persists and several trials shall be conducted to assess consistency.

show abstract

Section: Loss Functionsmentioning

confidence: 99%

Investigating Loss Functions for Segmenting and Detecting Ships on Sar Imagery

Lagahit,

Matsuoka

2024

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.

View full text Add to dashboard Cite

show abstract

“…Which provides strong intensity values during light detection and ranging (LiDAR) scanning, clearly distinguishing them from other features. Recently, this approach has extended to sparse point cloud-derived images from low-cost mobile LiDAR scanning (MLS) as an alternative to expensive mobile mapping systems (Lagahit and Matsuoka, 2023). This was done in an attempt to utilize low-cost LiDAR sensors onboard self-driving vehicles as a mobile mapping resource for updating and making digital twins or high-definition (HD) maps more dynamic.…”

Section: Introduction 11 Backgroundmentioning

confidence: 99%

“…Currently, there are already a multitude of CNN models with varying structures available, U-Net and Fast-SCNN to name a few (Poudel et al, 2019;Ronneberger et al, 2015). Both of these models have already demonstrated potential in extracting road markings from sparse MLS point cloud-derived images (Lagahit and Matsuoka, 2023). It is worth noting, however, that Fast-SCNN, which was built for real-time segmentation, has shown 15x quicker prediction speeds than U-Net.…”

Section: Introduction 11 Backgroundmentioning

confidence: 99%

See 1 more Smart Citation

Mfscnn: Appending a Masked Branch to Fast-SCNN to Improve Road Marking Extraction on Sparse MLS Point Cloud-Derived Images

Lagahit,

Matsuoka

2023

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.

View full text Add to dashboard Cite

Abstract. With the rise of self-driving cars, an increasing number of vehicles are equipped with low-cost light detection and ranging (LiDAR) sensors that could potentially serve as a massive mobile mapping resource, particularly for jobs that require multiple and frequent scanning, such as maintaining dynamic high-definition maps or digital twins. However, low-cost LiDAR sensors produce sparser point clouds during scanning which can make deep learning techniques for the automatic retrieval of features difficult like extracting road markings. In this work, we aim to improve the performance of a convolutional neural network (CNN) model for road marking extraction from sparse mobile LiDAR scanning (MLS) point cloud-derived images. We propose the modification of the Fast-SCNN model structure by adding a 2D convolution branch with masking in the feature fusion step: MFSCNN. To retain speed we only use MFSCNN to boost model training and still utilize Fast-SCNN for inference. Our results indicate potential, with a 4.6% increase in mean f1-score and an 8% decrease in uncertainty for the road marking class after multiple trials. Additionally, this research aims to support and increase research interest in lower-cost LiDARs for mobile mapping.

show abstract

“…Recent research has attempted to extract features from sparse point cloud-derived images using convolutional neural networks (CNN) generated by low-cost mobile light detection and ranging (LiDAR) scanning. One example is the extraction of road markings such as lane lines and crossing marks that return relatively high intensity values (Lagahit & Matsuoka, 2023). The successful extraction of features from sparse point clouds enables the usage of low-cost LiDARs which leads to a more practical alternative for mobile mapping tasks, especially for those that monitor and track changes in dynamic environments.…”

Section: Introduction 11 Backgroundmentioning

confidence: 99%

Exploring Ground Segmentation From Lidar Scanning-Derived Images Using Convolutional Neural Networks

Lagahit

Sakaguchi

et al. 2023

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.

View full text Add to dashboard Cite

Abstract. Recent works have attempted to extract features such as road markings from sparse mobile LiDAR scanning point cloud-derived images via convolutional neural networks (CNN). In this paper, the use of such methods for ground segmentation was explored. To begin, point clouds from each channel will be projected onto the y-z plane to generate the images that will be used for training and testing the CNN model. Then, for the main workflow, the following steps were performed for each channel: (1) point cloud-to-image conversion; (2) CNN classification; and (3) image-to-point cloud projection. Then utilizing multi-threading, each channel is processed in parallel to generate our ground-segmented sparse point cloud. Our findings have shown successful ground segmentation, achieving an f1-score of 98.9%. However, it performed 27.81% slower as compared to RANSAC. Overall, this initial investigation has demonstrated that ground segmentation from sparse point cloud-derived imagery is possible, and with further improvements to the CNN model, to make it faster, it has good potential to act as an alternative to conventional point cloud processing.

show abstract

Focal Combo Loss for Improved Road Marking Extraction of Sparse Mobile LiDAR Scanning Point Cloud-Derived Images Using Convolutional Neural Networks

Cited by 6 publications

References 29 publications

Investigating Loss Functions for Segmenting and Detecting Ships on Sar Imagery

Investigating Loss Functions for Segmenting and Detecting Ships on Sar Imagery

Mfscnn: Appending a Masked Branch to Fast-SCNN to Improve Road Marking Extraction on Sparse MLS Point Cloud-Derived Images

Exploring Ground Segmentation From Lidar Scanning-Derived Images Using Convolutional Neural Networks

Contact Info

Product

Resources

About