Start Date
5-12-2025 12:00 PM
End Date
5-12-2025 1:00 PM
Description
In many robotics or computer vision contained systems, regulating the quality of captured input data (e.g., point clouds or videos) is critical for preserving the integrity and accuracy of downstream tasks. In practice, this process requires mitigating various forms of noise specific to the domain (2D vs. 3D) and the intended application, such as object detection and scene understanding.
However, a unified framework addressing noise across a holistic system remains underexplored, as current surveys typically isolate specific modalities i.e., examine either 2D or 3D methods exclusively. The 2D domain covers two main areas: static image denoising, which prioritizes sensor noise modeling (e.g., Gaussian-Poisson mixtures) and feature preservation, and dynamic video denoising, which tackles motion blur and temporal consistency. Conversely, 3D LiDAR research remains largely distinct, focusing primarily on geometric challenges such as point sparsity, non-uniform density, and sensor drop-off. While effectively consolidating LiDAR-focused techniques and architectural trends, recent work by Wang et al. analyzes the field solely through supervision levels and modeling perspectives and leaves the integration with 2D modalities unaddressed.
Furthermore, the lack of a unified perspective creates a critical gap in the evaluation of emerging multi-modal systems. Modern architectures increasingly process 2D (RGB) and 3D (LiDAR/Depth) data simultaneously to leverage complementary information. However, diagnosing failures in these fusion models remains a challenge due to the opacity of current evaluation metrics. When a multi-modal system underperforms, it is often unclear whether the error stems from domain-specific degradation e.g., 2D motion blur or 3D point sparsity, or from the fusion process itself e.g., spatiotemporal misalignment. Without a framework to independently address these variables, there remains much struggle in the interpretability and robustness of multi-modal denoising systems.
We argue that 2D and 3D denoising techniques essentially address analogous problems in a visual contained system. For instance, “grainy” static images 1 (2D sensor noise) share fundamental characteristics with “noisy range measurements” in LiDAR (3D sensor noise). Similarly, “video flicker” (2D temporal noise) is mathematically analogous to “point cloud jitter” in dynamic SLAM (3D temporal noise). Thus, our main contribution is a universal four-domain evaluation framework across different domains to address visual noise from its four common origins: (1) Sensor, (2) Feature, (3) Temporal, and (4) System.
Our framework also offers a timely and comprehensive evaluation methodology for these complex systems. By disentangling noise sources into universal domains (Sensor, System, Temporal, Feature) rather than modality-specific evaluation techniques, researchers can validate a fusion algorithm's performance with greater granularity. For example, instead of simply reporting lower accuracy, a developer can use our framework to identify that a model handles 3D Feature noise well but fails catastrophically against 2D Sensor noise. This capability not only simplifies the benchmarking of multi-modal algorithms but also has the ability to transform a complex, entangled evaluation process into a structured, domain-specific checklist. In general, our framework aims to provide a consistent way to reason about noise sources, filtering strategies, and their roles in the overall visual contained systems, regardless of whether the data originates from a camera or from LiDAR.
A Universal Evaluation Framework for Visual Noise in Contained Systems
In many robotics or computer vision contained systems, regulating the quality of captured input data (e.g., point clouds or videos) is critical for preserving the integrity and accuracy of downstream tasks. In practice, this process requires mitigating various forms of noise specific to the domain (2D vs. 3D) and the intended application, such as object detection and scene understanding.
However, a unified framework addressing noise across a holistic system remains underexplored, as current surveys typically isolate specific modalities i.e., examine either 2D or 3D methods exclusively. The 2D domain covers two main areas: static image denoising, which prioritizes sensor noise modeling (e.g., Gaussian-Poisson mixtures) and feature preservation, and dynamic video denoising, which tackles motion blur and temporal consistency. Conversely, 3D LiDAR research remains largely distinct, focusing primarily on geometric challenges such as point sparsity, non-uniform density, and sensor drop-off. While effectively consolidating LiDAR-focused techniques and architectural trends, recent work by Wang et al. analyzes the field solely through supervision levels and modeling perspectives and leaves the integration with 2D modalities unaddressed.
Furthermore, the lack of a unified perspective creates a critical gap in the evaluation of emerging multi-modal systems. Modern architectures increasingly process 2D (RGB) and 3D (LiDAR/Depth) data simultaneously to leverage complementary information. However, diagnosing failures in these fusion models remains a challenge due to the opacity of current evaluation metrics. When a multi-modal system underperforms, it is often unclear whether the error stems from domain-specific degradation e.g., 2D motion blur or 3D point sparsity, or from the fusion process itself e.g., spatiotemporal misalignment. Without a framework to independently address these variables, there remains much struggle in the interpretability and robustness of multi-modal denoising systems.
We argue that 2D and 3D denoising techniques essentially address analogous problems in a visual contained system. For instance, “grainy” static images 1 (2D sensor noise) share fundamental characteristics with “noisy range measurements” in LiDAR (3D sensor noise). Similarly, “video flicker” (2D temporal noise) is mathematically analogous to “point cloud jitter” in dynamic SLAM (3D temporal noise). Thus, our main contribution is a universal four-domain evaluation framework across different domains to address visual noise from its four common origins: (1) Sensor, (2) Feature, (3) Temporal, and (4) System.
Our framework also offers a timely and comprehensive evaluation methodology for these complex systems. By disentangling noise sources into universal domains (Sensor, System, Temporal, Feature) rather than modality-specific evaluation techniques, researchers can validate a fusion algorithm's performance with greater granularity. For example, instead of simply reporting lower accuracy, a developer can use our framework to identify that a model handles 3D Feature noise well but fails catastrophically against 2D Sensor noise. This capability not only simplifies the benchmarking of multi-modal algorithms but also has the ability to transform a complex, entangled evaluation process into a structured, domain-specific checklist. In general, our framework aims to provide a consistent way to reason about noise sources, filtering strategies, and their roles in the overall visual contained systems, regardless of whether the data originates from a camera or from LiDAR.