EgoGaussian: Dynamic Scene Understanding from Egocentric Video with 3D Gaussian Splatting (2406.19811v2)

Abstract: Human activities are inherently complex, often involving numerous object interactions. To better understand these activities, it is crucial to model their interactions with the environment captured through dynamic changes. The recent availability of affordable head-mounted cameras and egocentric data offers a more accessible and efficient means to understand human-object interactions in 3D environments. However, most existing methods for human activity modeling neglect the dynamic interactions with objects, resulting in only static representations. The few existing solutions often require inputs from multiple sources, including multi-camera setups, depth-sensing cameras, or kinesthetic sensors. To this end, we introduce EgoGaussian, the first method capable of simultaneously reconstructing 3D scenes and dynamically tracking 3D object motion from RGB egocentric input alone. We leverage the uniquely discrete nature of Gaussian Splatting and segment dynamic interactions from the background, with both having explicit representations. Our approach employs a clip-level online learning pipeline that leverages the dynamic nature of human activities, allowing us to reconstruct the temporal evolution of the scene in chronological order and track rigid object motion. EgoGaussian shows significant improvements in terms of both dynamic object and background reconstruction quality compared to the state-of-the-art. We also qualitatively demonstrate the high quality of the reconstructed models.

• The paper demonstrates a novel approach using 3D Gaussian splatting to separate static scenes from dynamic object interactions using only RGB egocentric video.
• It employs a two-phase training strategy that first reconstructs the static background and then refines dynamic object poses using rigid estimation techniques.
• Experiments show significant improvements over state-of-the-art methods in SSIM, PSNR, and LPIPS metrics on in-the-wild egocentric datasets.

EgoGaussian: Dynamic Scene Understanding from Egocentric Video with 3D Gaussian Splatting

Introduction

The paper "EgoGaussian: Dynamic Scene Understanding from Egocentric Video with 3D Gaussian Splatting" presents a novel approach to reconstruct 3D scenes and track dynamic object interactions using RGB egocentric video input. Egocentric data has become more accessible with advancements in affordable head-mounted cameras, opening new opportunities for understanding complex human activities and object interactions in 3D environments. EgoGaussian takes a significant step forward by providing a method that relies solely on RGB input, in contrast to existing methods that often depend on multi-camera setups, depth-sensing cameras, or additional sensors.

Problem Statement

Human activities involve intricate interactions with multiple objects, and effectively modeling these dynamic interactions is crucial for understanding behavior. Traditional techniques either focus on reconstructing static 3D scenes or require extensive multi-source input to capture dynamics, leading to static representations fraught with artifacts like the "ghost effect." This paper introduces a method that overcomes these limitations by combining the strengths of egocentric video capture and 3D Gaussian Splatting.

Methodology

EgoGaussian builds on the framework of 3D Gaussian Splatting (3D-GS), explicitly using a set of 3D Gaussians characterized by position, covariance, opacity, and color features to represent the scene. The method identifies critical temporal points and partitions the video into static and dynamic clips. The static clips are used to reconstruct background scenes, whereas the dynamic clips capture object motion and refine object shapes.

Data Preprocessing

EgoGaussian uses an off-the-shelf method to obtain hand-object segmentation masks and derive camera poses through structure-from-motion (SfM). The video is partitioned into static and dynamic clips based on segmentation masks, with static clips reconstructing the background and dynamic clips focusing on object motion.

Static Clip Reconstruction

The initial training phase utilizes static clips to capture the background while dismissing dynamic objects to avoid inconsistencies. A binary cross-entropy loss is employed to differentiate background Gaussians from object Gaussians, allowing later object-specific refinements. This step is crucial for disentangling the static scene from dynamic interactions.

Dynamic Object Modeling

Dynamic clips introduce additional complexity as they involve object motion. EgoGaussian applies rigid object pose estimation techniques to track these movements across video frames. The training involves alternating phases of optimizing object poses and refining Gaussian parameters, leading to accurate dynamic object reconstructions that integrate seamlessly with the static background.

Evaluation and Results

The method is evaluated against existing state-of-the-art (SOTA) techniques like Deformable 3DGS and 4DGS using in-the-wild egocentric video datasets, HOI4D and EPIC-KITCHENS. The metrics used for evaluation include SSIM, PSNR, and LPIPS, focusing on the quality of reconstructions without actor influences. EgoGaussian outperforms the SOTA methods significantly, achieving better quantitative and qualitative results in both static and dynamic scenes.

Implications and Future Work

The practical implications of EgoGaussian are profound, facilitating detailed and accurate reconstructions of dynamic scenes from egocentric videos. This can enhance applications in behavioral analysis, augmented reality, and robotics, where understanding object interactions is critical. Theoretically, the method contributes to the field of dynamic scene understanding, setting a new benchmark for using RGB input to capture and reconstruct complex interactions.

Future Developments

While EgoGaussian effectively handles rigid objects, future work could extend its capabilities to model elastic or stretchable objects, further broadening its application scope. Additionally, optimizing the training time and refining background-object integration can enhance the method's efficiency and accuracy.

Conclusion

EgoGaussian introduces an innovative method for dynamic scene understanding, leveraging 3D Gaussian Splatting from RGB egocentric video alone. By outperforming existing methods in both static and dynamic settings, it opens new avenues for 3D scene reconstruction and dynamic interaction modeling. Future enhancements could further elevate its applicability and impact across various domains in artificial intelligence and computer vision.