A Review of panoptic segmentation for mobile mapping point clouds DOI Creative Commons
Binbin Xiang, Yuanwen Yue, Torben Peters

et al.

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 203, P. 373 - 391

Published: Aug. 31, 2023

3D point cloud panoptic segmentation is the combined task to (i) assign each a semantic class and (ii) separate points in into object instances. Recently there has been an increased interest such comprehensive scene understanding, building on rapid advances of due advent deep neural networks. Yet, date very little work about outdoor mobile-mapping data, no systematic comparisons. The present paper tries close that gap. It reviews blocks needed assemble pipeline related literature. Moreover, modular set up perform comprehensive, experiments assess state context street mapping. As byproduct, we also provide first public dataset for task, by extending NPM3D include instance labels. That our source code are publicly available.1We discuss which adaptations need adapt current methods scenes large objects. Our study finds mobile mapping KPConv performs best but slower, while PointNet++ fastest significantly worse. Sparse CNNs between. Regardless backbone, clustering embedding features better than using shifted coordinates.

Language: Английский

A Survey on Deep-Learning-Based LiDAR 3D Object Detection for Autonomous Driving DOI Creative Commons
Simegnew Yihunie Alaba, John Ball

Sensors, Journal Year: 2022, Volume and Issue: 22(24), P. 9577 - 9577

Published: Dec. 7, 2022

LiDAR is a commonly used sensor for autonomous driving to make accurate, robust, and fast decision-making when driving. The in the perception system, especially object detection, understand environment. Although 2D detection has succeeded during deep-learning era, lack of depth information limits understanding environment location. Three-dimensional sensors, such as LiDAR, give 3D about surrounding environment, which essential system. Despite attention computer vision community due multiple applications robotics driving, there are challenges, scale change, sparsity, uneven distribution data, occlusions. Different representations data methods minimize effect sparsity have been proposed. This survey presents LiDAR-based feature-extraction techniques data. coordinate systems differ camera datasets methods. Therefore, summarized. Then, state-of-the-art object-detection reviewed with selected comparison among

Language: Английский

Citations

57
Class-Aware Fish Species Recognition Using Deep Learning for an Imbalanced Dataset DOI Creative Commons
Simegnew Yihunie Alaba, M M Nabi, Chiranjibi Shah

et al.

Sensors, Journal Year: 2022, Volume and Issue: 22(21), P. 8268 - 8268

Published: Oct. 28, 2022

Fish species recognition is crucial to identifying the abundance of fish in a specific area, controlling production management, and monitoring ecosystem, especially endangered species, which makes accurate essential. In this work, problem formulated as an object detection model handle multiple single image, challenging classify using simple classification network. The proposed consists MobileNetv3-large VGG16 backbone networks SSD head. Moreover, class-aware loss function solve class imbalance our dataset. takes number instances each into account gives more weight those with smaller instances. This can be applied any or task imbalanced experimental result on large-scale reef dataset, SEAMAPD21, shows that improves over original by up 79.7%. Pascal VOC dataset also outperforms model.

Language: Английский

Citations

44
Deep Learning-Based Image 3-D Object Detection for Autonomous Driving: Review DOI
Simegnew Yihunie Alaba, John Ball

IEEE Sensors Journal, Journal Year: 2023, Volume and Issue: 23(4), P. 3378 - 3394

Published: Jan. 13, 2023

An accurate and robust perception system is key to understanding the driving environment of autonomous robots. Autonomous needs 3-D information about objects, including object’s location pose, understand clearly. A camera sensor widely used in because its richness color texture, low price. The major problem with lack information, which necessary environment. In addition, scale change occlusion make object detection more challenging. Many deep learning-based methods, such as depth estimation, have been developed solve information. This survey presents image bounding box encoding techniques evaluation metrics. image-based methods are categorized based on technique estimate an image’s insights added each method. Then, state-of-the-art (SOTA) monocular stereo camera-based summarized. We also compare performance selected models present challenges future directions detection.

Language: Английский

Citations

42
Classification of Lung Diseases Using an Attention-Based Modified DenseNet Model DOI

Upasana Chutia,

Anand Shanker Tewari, Jyoti Prakash Singh

et al.

Deleted Journal, Journal Year: 2024, Volume and Issue: 37(4), P. 1625 - 1641

Published: March 11, 2024

Lung diseases represent a significant global health threat, impacting both well-being and mortality rates. Diagnostic procedures such as Computed Tomography (CT) scans X-ray imaging play pivotal role in identifying these conditions. X-rays, due to their easy accessibility affordability, serve convenient cost-effective option for diagnosing lung diseases. Our proposed method utilized the Contrast-Limited Adaptive Histogram Equalization (CLAHE) enhancement technique on images highlight key feature maps related using DenseNet201. We have augmented existing Densenet201 model with hybrid pooling channel attention mechanism. The experimental results demonstrate superiority of our over well-known pre-trained models, VGG16, VGG19, InceptionV3, Xception, ResNet50, ResNet152, ResNet50V2, ResNet152V2, MobileNetV2, DenseNet121, DenseNet169, achieves impressive accuracy, precision, recall, F1-scores 95.34%, 97%, 96%, respectively. also provide visual insights into model's decision-making process Gradient-weighted Class Activation Mapping (Grad-CAM) identify normal, pneumothorax, atelectasis cases. terms heatmap may help radiologists improve diagnostic abilities labelling processes.

Language: Английский

Citations

11
Off-Road Detection Analysis for Autonomous Ground Vehicles: A Review DOI Creative Commons
Fahmida Islam, M M Nabi, John Ball

et al.

Sensors, Journal Year: 2022, Volume and Issue: 22(21), P. 8463 - 8463

Published: Nov. 3, 2022

When it comes to some essential abilities of autonomous ground vehicles (AGV), detection is one them. In order safely navigate through any known or unknown environment, AGV must be able detect important elements on the path. Detection applicable both on-road and off-road, but they are much different in each environment. The key environment that identify drivable pathway whether there obstacles around it. Many works have been published focusing components various ways. this paper, a survey most recent advancements methods intended specifically for off-road has presented. For this, we divided literature into three major groups: positive negative obstacles. Each portion further multiple categories based technology used, example, single sensor-based, how data analyzed. Furthermore, added critical findings technology, challenges associated with possible future directions. Authors believe work will help reader finding who doing similar works.

Language: Английский

Citations

29
WCNN3D: Wavelet Convolutional Neural Network-Based 3D Object Detection for Autonomous Driving DOI Creative Commons
Simegnew Yihunie Alaba, John Ball

Sensors, Journal Year: 2022, Volume and Issue: 22(18), P. 7010 - 7010

Published: Sept. 16, 2022

Three-dimensional object detection is crucial for autonomous driving to understand the environment. Since pooling operation causes information loss in standard CNN, we designed a wavelet-multiresolution-analysis-based 3D network without operation. Additionally, instead of using single filter like convolution, used lower-frequency and higher-frequency coefficients as filter. These filters capture more relevant parts than filter, enlarging receptive field. The model comprises discrete wavelet transform (DWT) an inverse (IWT) with skip connections encourage feature reuse contrasting expanding layers. IWT enriches representation by fully recovering lost details during downsampling Element-wise summation was decrease computational burden. We trained Haar Daubechies (Db4) wavelets. two-level decomposition result shows that can build lightweight losing significant performance. experimental results on KITTI's BEV evaluation benchmark show our outperforms PointPillars-based up 14% while reducing number trainable parameters.

Language: Английский

Citations

24
Benchmarking 2D Multi-Object Detection and Tracking Algorithms in Autonomous Vehicle Driving Scenarios DOI Creative Commons
Diego Gragnaniello, Antonio Greco, Alessia Saggese

et al.

Sensors, Journal Year: 2023, Volume and Issue: 23(8), P. 4024 - 4024

Published: April 16, 2023

Self-driving vehicles must be controlled by navigation algorithms that ensure safe driving for passengers, pedestrians and other vehicle drivers. One of the key factors to achieve this goal is availability effective multi-object detection tracking algorithms, which allow estimate position, orientation speed on road. The experimental analyses conducted so far have not thoroughly evaluated effectiveness these methods in road scenarios. To aim, we propose paper a benchmark modern applied image sequences acquired camera installed board vehicle, namely, videos available BDD100K dataset. proposed framework allows evaluate 22 different combinations using metrics highlight positive contribution limitations each module considered algorithms. analysis results points out best method currently combination ConvNext QDTrack, but also images substantially improved. Thanks our analysis, conclude evaluation should extended considering specific aspects autonomous scenarios, such as multi-class problem formulation distance from targets, simulating impact errors safety.

Language: Английский

Citations

14
Deep learning based computer vision under the prism of 3D point clouds: a systematic review DOI Creative Commons
Kyriaki A. Tychola, Εleni Vrochidou, George A. Papakostas

et al.

The Visual Computer, Journal Year: 2024, Volume and Issue: unknown

Published: Jan. 29, 2024

Abstract Point clouds consist of 3D data points and are among the most considerable formats for representations. Their popularity is due to their broad application areas, such as robotics autonomous driving, employment in basic vision tasks segmentation, classification, detection. However, processing point challenging compared other visual forms images, mainly unstructured nature. Deep learning (DL) has been established a powerful tool processing, reporting remarkable performance enhancements traditional methods all 2D tasks. However new challenges emerging when it comes clouds. This work aims guide future research by providing systematic review DL on clouds, holistically covering technologies cloud formation reviewed each other. The discussed, state-of-the-art models’ performances focusing solutions. Moreover, this popular benchmark datasets summarized based task-oriented applications, aiming highlight existing constraints comparatively evaluate them. Future directions upcoming trends also highlighted.

Language: Английский

Citations

5
Scalable 3D Panoptic Segmentation As Superpoint Graph Clustering DOI
Damien Robert, Hugo Raguet, Loïc Landrieu

et al.

2021 International Conference on 3D Vision (3DV), Journal Year: 2024, Volume and Issue: unknown, P. 179 - 189

Published: March 18, 2024

We introduce a highly efficient method for panoptic segmentation of large 3D point clouds by redefining this task as scalable graph clustering problem. This approach can be trained using only local auxiliary tasks, thereby eliminating the resource-intensive instance-matching step during training. Moreover, our formulation easily adapted to superpoint paradigm, further increasing its efficiency. allows model process scenes with millions points and thousands objects in single inference. Our method, called SuperCluster, achieves new state-of-the-art performance two indoor scanning datasets: 50.1 PQ (+7.8) S3DIS Area 5, 58.7 (+25.2) ScanNetV2. also set first large-scale mobile mapping benchmarks: KITTI-360 DALES. With 209k parameters, is over 30 times smaller than best-competing trains up 15 faster. code pretrained models are available at https://github.com/drprojects/superpoint_transformer.

Language: Английский

Citations

5
Deep learning for 3D object recognition: A survey DOI
A. A. M. Muzahid, Hua Han,

Yujin Zhang

et al.

Neurocomputing, Journal Year: 2024, Volume and Issue: 608, P. 128436 - 128436

Published: Aug. 23, 2024

Language: Английский

Citations

4