In the realm of 3D-computer vision applications, point cloud few-shot learning plays a critical role. However, it poses an arduous challenge due to the sparsity, irregularity, and unordered nature of the data. Current methods rely on complex local geometric extraction techniques such as convolution, graph, and attention mechanisms, along with extensive data-driven pre-training tasks. These approaches contradict the fundamental goal of few-shot learning, which is to facilitate efficient learning. To address this issue, we propose GPr-Net (Geometric Prototypical Network), a lightweight and computationally efficient geometric prototypical network that captures the intrinsic topology of point clouds and achieves superior performance. Our proposed method, IGI++(Intrinsic Geometry Interpreter++) employs vector-based hand-crafted intrinsic geometry interpreters and Laplace vectors to extract and evaluate point cloud morphology, resulting in improved representations for FSL (Few-Shot Learning). Additionally, Laplace vectors enable the extraction of valuable features from point clouds with fewer points. To tackle the distribution drift challenge in few-shot metric learning, we leverage hyperbolic space and demonstrate that our approach handles intra and inter-class variance better than existing point cloud few-shot learning methods. Experimental results on the ModelNet40 dataset show that GPr-Net outperforms state-of-the-art methods in few-shot learning on point clouds, achieving utmost computational efficiency that is 170x better than all existing works. The code is publicly available at https://github. com/TejasAnvekar/GPr-Net.
CVPR-W
IPD-Net: SO(3) Invariant Primitive Decompositional Network for 3D Point Clouds
Ramesh Ashok Tabib, Nitishkumar Upasi, Tejas Anvekar, Dikshit Hegde, and Uma Mudenagudi
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Jun 2023
In this paper, we propose IPD-Net: Invariant Primitive Decompositional Network, a SO (3) invariant framework for decomposition of a point cloud. The human cognitive system is able to identify and interpret familiar objects regardless of their orientation and abstraction. Recent research aims to bring this capability to machines for understanding the 3D world. In this work, we present a framework inspired by human cognition to decompose point clouds into four primitive 3D shapes (plane, cylinder, cone, and sphere) and enable machines to understand the objects in various orientations. We employ Implicit Invariant Features (IIF) to learn local geometric relations by implicitly representing the point cloud with enhanced geometric information invariant towards SO (3) rotations. We also use Spatial Rectification Unit (SRU) to extract invariant global signatures. We demonstrate the results of our proposed methodology for SO (3) invariant decomposition on TraceParts Dataset, and show the generalizability of proposed IPD-Net as plugin for downstream task on classification of point clouds. We compare the results of classification with state-of-the-art methods on benchmark dataset (ModelNet40).
PointCLIMB: An Exemplar-Free Point Cloud Class Incremental Benchmark
Shivanand Kundargi, Tejas Anvekar, Ramesh Ashok Tabib, and Uma Mudenagudi
Point clouds offer comprehensive and precise data regarding the contour and configuration of objects. Employing such geometric and topological 3D information of objects in class incremental learning can aid endless application in 3D-computer vision. Well known 3D-point cloud class incremental learning methods for addressing catastrophic forgetting generally entail the usage of previously encountered data, which can present difficulties in situations where there are restrictions on memory or when there are concerns about the legality of the data. Towards this we pioneer to leverage exemplar free class incremental learning on Point Clouds. In this paper we propose PointCLIMB: An exemplar Free Class Incremental Learning Benchmark. We focus on a pragmatic perspective to consider novel classes for class incremental learning on 3D point clouds. We setup a benchmark for 3D Exemplar free class incremental learning. We investigate performance of various backbones on 3D-Exemplar Free Class Incremental Learning framework. We demonstrate our results on ModelNet40 dataset.
WACV-C
1st Workshop on Maritime Computer Vision (MaCVi) 2023: Challenge Results
Benjamin Kiefer, Matej Kristan, Janez Perš, Lojze Žust, Fabio Poiesi, and 67 more authors
In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, Jan 2023
The 1st Workshop on Maritime Computer Vision (MaCVi)| 2023 focused on maritime computer vision for Unmanned| Aerial Vehicles (UAV) and Unmanned Surface Vehicle (USV),| and organized several subchallenges in this domain:(i) UAV-based Maritime Object Detection,(ii) UAV-based Maritime Object Tracking,(iii) USV-based Maritime Obstacle Segmentation and (iv) USV-based Maritime Obstacle Detection. The subchallenges were based on the SeaDronesSee and MODS| benchmarks. This report summarizes the main findings of the| individual subchallenges and introduces a new benchmark,| called SeaDronesSee Object Detection v2, which extends the| previous benchmark by including more classes and footage.| We provide statistical and qualitative analyses, and assess| trends in the best-performing methodologies of over 130| submissions. The methods are summarized in the appendix.| The datasets, evaluation code and the leaderboard are| publicly available (https://seadronessee. cs. uni-tuebingen. de/macvi).
ICCV-W
ASUR3D: Arbitrary Scale Upsampling and Refinement of 3D Point Clouds Using Local Occupancy Fields
Akash Kumbar, Tejas Anvekar, Ramesh Ashok Tabib, and Uma Mudenagudi
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2023
In this paper, we introduce ASUR3D, a novel methodology for the arbitrary-scale upsampling of 3D point clouds employing Local Occupancy Representation. Our proposed implicit occupancy representation enables efficient point classification, effectively discerning points belonging to the surface from non-surface points. Learning an implicit representation of open surfaces, enables one to capture the better local neighbourhood representation, leading to finer refinement and reconstruction with enhanced preservation of intricate geometric details. Leveraging this capability, we can accurately sample an arbitrary number of points on the surface, facilitating precise and flexible upsampling. We demonstrate the effectiveness of ASUR3D on PUGAN and PU1K benchmark datasets. Our proposed method achieves state-of-the-art results on all benchmarks and for all evaluation metrics. Additionally, we demonstrate the efficacy of our methodology on self-proposed heritage data generated through photogrammetry, further confirming its effectiveness in diverse scenarios. The code is publicly available at https://github. com/Akash-Kumbar/ASUR3D
ICCV-W
DeFi: Detection and Filling of Holes in Point Clouds Towards Restoration of Digitized Cultural Heritage Models
Ramesh Ashok Tabib, Dikshit Hegde, Tejas Anvekar, and Uma Mudenagudi
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2023
In this paper, we propose DeFi: a novel perspective for hole detection and filling of a given deteriorated 3D point cloud towards digital preservation of cultural heritage sites. Preservation of heritage demands digitization as cultural heritage sites deteriorate due to natural calamities and human activities. Digital preservation promotes acquisition of 3D data using 3D sensor or Multi-view reconstruction. Unfortunately, 3D data acquisition finds challenges due to the limitations in sensor technology and inappropriate capture conditions, leading to formation of missing regions or holes in the acquired point cloud. To address this, we propose a pipeline consisting of detection of hole boundaries, and understanding the geometry of the hole boundaries to fill the region of the point cloud. Recent research on hole detection and filling fails to generalize on complex structures such as heritage sites, as they find challenges in differentiating between the hole boundary and non-hole boundary points. To address this, we propose to detect boundary points of point cloud and learn to classify them into" hole boundary" and" non-hole boundary" points. We generate a synthetic dataset based on ModelNet40 to learn the detection of hole boundaries. We demonstrate the results of the proposed pipeline on (i) ModelNet40 dataset,(ii) Heritage 3D models generated via photogrammetry, and compare the results with state-of-the-art methods.
ICCV-W
TP-NoDe: Topology-Aware Progressive Noising and Denoising of Point Clouds Towards Upsampling
Akash Kumbar, Tejas Anvekar, Tulasi Amitha Vikrama, Ramesh Ashok Tabib, and Uma Mudenagudi
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2023
In this paper, we propose TP-NoDe, a novel Topology-aware Progressive Noising and Denoising technique for 3D point cloud upsampling. TP-NoDe revisits the traditional method of upsampling of the point cloud by introducing a novel perspective of adding local topological noise by incorporating a novel algorithm Density-Aware k nearest neighbour (DA-kNN) followed by denoising to map noisy perturbations to the topology of the point cloud. Unlike previous methods, we progressively upsample the point cloud, starting at a 2 x upsampling ratio and advancing to a desired ratio. TP-NoDe generates intermediate upsampling resolutions for free, obviating the need to train different models for varying upsampling ratios. TP-NoDe mitigates the need for task-specific training of upsampling networks for a specific upsampling ratio by reusing a point cloud denoising framework. We demonstrate the supremacy of our method TP-NoDe on the PU-GAN dataset and compare it with state-of-the-art upsampling methods. The code is publicly available at https://github. com/Akash-Kumbar/TPNoDe.
2022
CVPR-W
DA-AE: Disparity-Alleviation Auto-Encoder Towards Categorization of Heritage Images for Aggrandized 3D Reconstruction.
Dikshit Hegde, Tejas Anvekar, Ramesh Ashok Tabib, and Uma Mudengudi
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Jun 2022
In this paper, we propose DA-AE: Disparity Alleviation AutoEncoder for categorization of heritage images towards 3D reconstruction. Recent survey on preservation of heritage shows demand for the digitization and conservations of heritage sites owing to their susceptibility to natural disasters and human acts. Digital conservation can be facilitated via crowdsourcing of data useful for construction of 3D models. Data from multiple sites sourced may result in elimination of relevant images due to the limitations of the pipeline. Curation and categorization of the crowdsourced data enables better 3D reconstruction. 3D reconstruction pipelines demand correlation between the data and also tries to eliminate the irrelevant information. The reconstruction pipeline is sensitive to selection of initial pair for reconstruction. By categorising individual sites, crowdsourced data can be used to create better 3D reconstructed models. Categorization of crowdsourced data demands learning robust representations of data. Towards this, we propose DA-AE for improved representation and categorization of data in latent space, along with a disparity alleviation loss. We demonstrate categorization as an event, with clustering as a downstream task. We compare our results of clustering with state-of-the-art methods on benchmark datasets (MNIST, FashionMNIST, and USPS). We demonstrate the effects of our categorization using custom dataset IDH10 and compare the results with state-of-the-art methods. We show a systematic and qualitative influence of the proposed method on 3D reconstruction of data.
SIGGRAPH-ASIA
Metric-KNN is All You Need
Tejas Anvekar, Ramesh Ashok Tabib, Dikshit Hegde, and Uma Mudenagudi
In this paper, we propose VG-VAE: Venatus Geometric Variational Auto-Encoder for capturing unsupervised hierarchical local and global geometric signatures in pointcloud. Recent research emphasises the significance of the underlying intrinsic geometry for pointcloud processing. Our contribution is to extract and analyse the morphology of the pointcloud using the proposed Geometric Proximity Correlator (GPC) and variational sampling of the latent. The extraction of local geometric signatures is facilitated by the GPC, whereas the extraction of global geometry is facilitated by variational sampling. Furthermore, we apply a naive mix of vector algebra and 3D geometry to extract the basic per-point geometric signature, which assists the unsupervised hypothesis. We provide statistical analyses of local and global geometric signatures. The impacts of our geometric features are demonstrated on pointcloud classification as downstream task using the classic pointcloud feature extractor PointNet. We demonstrate our analysis on ModelNet40 a benchmark dataset, and compare with state-of-the-art techniques.