TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction

arxiv: 2604.02586 · v1 · submitted 2026-04-02 · 💻 cs.CV · cs.GR

TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction

Daheng Yin , Isaac Ding , Yili Jin , Jianxin Shi , Jiangchuan Liu This is my paper

Pith reviewed 2026-05-13 20:53 UTC · model grok-4.3

classification 💻 cs.CV cs.GR

keywords 3D Gaussian SplattingDynamic Scene ReconstructionPoint TrackingMotion CompensationVideo-based 3D ReconstructionParallel OptimizationTemporal Consistency

0 comments p. Extension

The pith

TrackerSplat uses 2D point tracks to reposition 3D Gaussians before optimization, handling large inter-frame motions without fading artifacts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a method that first runs off-the-shelf point trackers on video frames to recover pixel trajectories. These 2D paths are triangulated into 3D and used to move, rotate, and resize individual Gaussians so they start near their correct locations for the next frame. Once the Gaussians are pre-aligned this way, standard gradient optimization can finish the reconstruction without the large search space that normally produces color shifts and disappearances under fast motion. The same pre-alignment step also lets the system train several adjacent frames at once on separate devices, raising throughput while the final rendered quality stays comparable to slower sequential baselines.

Core claim

TrackerSplat extracts per-pixel trajectories with existing point trackers, triangulates those trajectories across views to obtain 3D displacements, and applies the resulting transformations to relocate, reorient, and rescale each Gaussian primitive before the usual training loop begins.

What carries the argument

Triangulation of 2D point trajectories onto 3D Gaussians that then supplies explicit relocation, rotation, and scaling updates prior to gradient optimization.

If this is right

Large inter-frame displacements no longer force sequential frame-by-frame processing.
Parallel training of multiple frames across devices becomes feasible without quality drop.
Fading and recoloring artifacts drop markedly on real-world fast-motion footage.
The same pre-positioning step works for any number of adjacent frames provided the trackers succeed.
Rendering quality remains comparable to sequential baselines on the tested real-world datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same pre-alignment idea could be tested on other explicit 3D representations such as point clouds or meshes to see whether motion handling separates cleanly from later refinement.
In robotics applications the approach might allow incremental updates of dynamic environments at higher frame rates than current Gaussian pipelines.
Replacing the off-the-shelf tracker with a learned one tuned for the reconstruction task could further reduce sensitivity to lighting changes or repetitive textures.

Load-bearing premise

Off-the-shelf point trackers produce 2D trajectories accurate and consistent enough across views to yield reliable 3D guidance for Gaussian placement even when objects move quickly or are partly occluded.

What would settle it

Run the method on a video sequence where a standard point tracker loses lock on fast-moving objects; the resulting 3D reconstruction should then exhibit the same fading and recoloring artifacts that appear in prior Gaussian methods without pre-alignment.

Figures

Figures reproduced from arXiv: 2604.02586 by Daheng Yin, Isaac Ding, Jiangchuan Liu, Jianxin Shi, Yili Jin.

**Figure 1.** Figure 1: Illustration of the motivation and basic idea of TrackerSplat. (a) Ground truth from the "walking" sequence. (b), (c) Rendered frames 64 and 67, trained [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: DOT point tracking on a video sequence. Colored lines show pixel [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: TrackerSplat overview. Our method processes video clips captured from multiple fixed viewpoints. It begins by applying existing reconstruction [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: TrackerSplat parallel pipeline. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Average visual quality (PSNR ↑ / SSIM ↑ / LPIPs ↓) over long-video sequences using our parallel pipeline with 8 GPUs (long-video experiments). Our method achieves higher and more stable visual quality than baselines in most cases, demonstrating its robustness. Lines ending prematurely for 4DGS and ST-4DGS indicate training failures due to GPU memory overflow (exceeding the 40GB limit of the A100 GPU) or nu… view at source ↗

**Figure 6.** Figure 6: Qualitative comparison of rendered results from the final frame of representative 9-frame clips processed in parallel using 8 GPUs (short-clip [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

read the original abstract

Recent advancements in 3D Gaussian Splatting (3DGS) have demonstrated its potential for efficient and photorealistic 3D reconstructions, which is crucial for diverse applications such as robotics and immersive media. However, current Gaussian-based methods for dynamic scene reconstruction struggle with large inter-frame displacements, leading to artifacts and temporal inconsistencies under fast object motions. To address this, we introduce \textit{TrackerSplat}, a novel method that integrates advanced point tracking methods to enhance the robustness and scalability of 3DGS for dynamic scene reconstruction. TrackerSplat utilizes off-the-shelf point tracking models to extract pixel trajectories and triangulate per-view pixel trajectories onto 3D Gaussians to guide the relocation, rotation, and scaling of Gaussians before training. This strategy effectively handles large displacements between frames, dramatically reducing the fading and recoloring artifacts prevalent in prior methods. By accurately positioning Gaussians prior to gradient-based optimization, TrackerSplat overcomes the quality degradation associated with large frame gaps when processing multiple adjacent frames in parallel across multiple devices, thereby boosting reconstruction throughput while preserving rendering quality. Experiments on real-world datasets confirm the robustness of TrackerSplat in challenging scenarios with significant displacements, achieving superior throughput under parallel settings and maintaining visual quality compared to baselines. The code is available at https://github.com/yindaheng98/TrackerSplat.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TrackerSplat pre-positions 3D Gaussians using triangulated point tracks from off-the-shelf models to cut artifacts from large motions, which is a practical engineering step but rests on untested tracker reliability.

read the letter

TrackerSplat's core move is to pull 2D trajectories from existing point trackers, triangulate them into 3D, and use those to shift, rotate, and scale the Gaussians before the standard optimization loop. This targets the fading and recoloring that show up when frames have big displacements, and it lets the pipeline handle multiple frames in parallel across devices without as much quality drop. The integration itself is not in the dynamic 3DGS papers cited in the abstract, so that specific pre-alignment step counts as new. It also keeps the method simple by leaning on mature trackers instead of building new ones from scratch. That focus on throughput while claiming to hold visual quality is the part that could matter for robotics or immersive pipelines that already run 3DGS. The experiments are described as confirming robustness on real-world data, which at least shows the authors tested the idea beyond toy cases. The main soft spot is that the whole pipeline inherits whatever errors the chosen trackers make on fast motion, occlusions, or lighting changes. The abstract gives no numbers on endpoint error, cross-view consistency, or what happens when tracks break, so the claim of dramatically fewer artifacts stays plausible but unverified in detail. If the trackers drift, the pre-adjustment could add new problems rather than remove them. No ablations on tracker choice or failure modes appear in the provided text either. This paper is for groups already working with dynamic 3D Gaussian pipelines who need faster processing on real scenes. It deserves a serious referee because the mechanism is concrete and the problem it attacks is common, even though the evidence will need tightening on the tracker side before it can be trusted at scale.

Referee Report

2 major / 1 minor

Summary. The paper introduces TrackerSplat, a method for dynamic 3D Gaussian Splatting reconstruction that uses off-the-shelf point tracking models to extract pixel trajectories, triangulates them into 3D, and employs the resulting 3D displacements to pre-adjust Gaussian positions, rotations, and scales before gradient-based optimization. This is claimed to mitigate artifacts from large inter-frame motions, reduce fading and recoloring issues common in prior dynamic 3DGS approaches, and enable parallel processing of adjacent frames across devices to increase throughput while preserving rendering quality. Experiments on real-world datasets are stated to confirm robustness in challenging scenarios with significant displacements and superior performance relative to baselines.

Significance. If the central claims hold under quantitative scrutiny, the work would be significant for dynamic scene reconstruction, as it offers a practical way to improve initialization robustness in 3DGS without inventing new trackers, directly addressing a known limitation for fast motions in robotics and immersive applications. The provision of reproducible code at the GitHub repository is a clear strength that facilitates verification and extension. The parallel-processing angle for throughput gains could be impactful if quality is demonstrably maintained.

major comments (2)

[Abstract and Experiments] Abstract and Experiments section: The robustness claim that the method 'dramatically reducing the fading and recoloring artifacts' and handles 'challenging scenarios with significant displacements' rests on the assumption that off-the-shelf trackers yield accurate, consistent, triangulable 2D trajectories. No quantitative metrics (e.g., endpoint error, occlusion failure rates, cross-view consistency) or ablation studies on tracker performance under fast motion, occlusions, or lighting changes are reported, making it impossible to assess whether the pre-adjustment step reliably outperforms standard 3DGS initialization.
[Method] Method section (trajectory triangulation and Gaussian guidance): The description of triangulating per-view 2D trajectories onto 3D Gaussians to guide relocation/rotation/scaling lacks detail on how inconsistencies across views or tracker failures are handled (e.g., outlier rejection, confidence weighting). This is load-bearing for the claim of artifact reduction, as any error in the 3D guidance directly propagates to the subsequent optimization.

minor comments (1)

[Abstract] The abstract provides a GitHub link for code, which supports reproducibility; ensure the released code includes the exact tracker configurations and parallelization scripts used in the reported experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address each major comment point by point below, providing clarifications and indicating planned changes to the manuscript.

read point-by-point responses

Referee: [Abstract and Experiments] The robustness claim that the method dramatically reduces fading and recoloring artifacts and handles challenging scenarios with significant displacements rests on the assumption that off-the-shelf trackers yield accurate, consistent, triangulable 2D trajectories. No quantitative metrics (e.g., endpoint error, occlusion failure rates) or ablation studies on tracker performance are reported.

Authors: We appreciate the referee highlighting this point. Our contribution centers on using established point trackers to pre-position Gaussians for improved 3DGS optimization in dynamic scenes, with end-to-end experiments showing reduced artifacts and better quality versus baselines on real-world data. We rely on the trackers' published performance (e.g., from their original papers) rather than re-evaluating them, as the focus is on the integration and its effect on reconstruction. To strengthen the manuscript, we will add a dedicated paragraph in the Experiments section discussing tracker reliability under the tested conditions and referencing their reported metrics, while noting that full tracker ablations fall outside the paper's scope. revision: partial
Referee: [Method] The description of triangulating per-view 2D trajectories onto 3D Gaussians to guide relocation/rotation/scaling lacks detail on how inconsistencies across views or tracker failures are handled (e.g., outlier rejection, confidence weighting).

Authors: We agree this requires clarification. Our pipeline applies RANSAC-based outlier rejection during multi-view triangulation to handle cross-view inconsistencies, and we filter trajectories using the tracker's per-point confidence scores, discarding those below a threshold or with high 3D projection variance. We will expand the Method section with explicit steps, pseudocode, and a supplementary figure detailing the guidance computation to make these mechanisms transparent. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method integrates external trackers with standard 3DGS optimization without self-referential reductions

full rationale

The paper's derivation relies on off-the-shelf point tracking models to extract 2D trajectories, followed by triangulation to 3D for pre-adjusting Gaussian parameters before gradient-based optimization. No equations or steps in the provided text reduce any claimed prediction or result to a fitted parameter or self-citation by construction. The approach treats trackers as independent external inputs and uses standard geometric triangulation, keeping the central mechanism self-contained against external benchmarks rather than circular.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on the standard assumptions of 3D Gaussian Splatting and the reliability of existing point trackers; no new free parameters, axioms, or invented entities are introduced in the abstract description.

axioms (2)

domain assumption 3D Gaussians can be optimized via gradient descent to represent scene appearance and geometry
Core premise inherited from prior 3DGS work and invoked when the method states that pre-positioning precedes training.
domain assumption Off-the-shelf point trackers produce trajectories accurate enough for 3D triangulation in dynamic real-world scenes
Directly required for the triangulation step that guides Gaussian relocation.

pith-pipeline@v0.9.0 · 5555 in / 1353 out tokens · 51507 ms · 2026-05-13T20:53:25.714651+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

[1]

InACM SIGGRAPH 2024 Conference Papers (SIGGRAPH ’24)

4D-Rotor Gaussian Splatting: Towards Efficient Novel View Synthesis for Dynamic Scenes. InACM SIGGRAPH 2024 Conference Papers (SIGGRAPH ’24). 1–11. Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang

work page 2024
[2]

Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds

InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds. doi:10.48550/ARXIV.2403.20309 Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu

work page doi:10.48550/arxiv.2403.20309
[3]

Taichi: a language for high-performance computation on spatially sparse data structures.ACM Transactions on Graphics (TOG)38, 6 (2019),

work page 2019
[4]

Co- tracker3: Simpler and better point tracking by pseudo- labelling real videos

CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos. doi:10.48550/ARXIV.2410.11831 SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong. TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction•9 Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi,...

work page doi:10.48550/arxiv.2410.11831 2025
[5]

InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.)

CoTracker: It Is Better to Track Together. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). 18–35. Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis

work page 2024
[6]

Guillaume Le Moing, Jean Ponce, and Cordelia Schmid

3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (2023), 139:1–139:14. Guillaume Le Moing, Jean Ponce, and Cordelia Schmid

work page 2023
[7]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. 2024b. ST-4DGS: Spatial-Temporally Consistent 4D Gaussian Splatting for Efficient Dynamic Scene Rendering. InACM SIGGRAPH 2024 Con...

work page 2024
[8]

arXiv preprint arXiv:2409.02104 , year=

DynOMo: Online Point Tracking by Dynamic Online Monocular Gaussian Reconstruction. arXiv:2409.02104 [cs] doi:10.48550/arXiv.2409.02104 Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas

work page doi:10.48550/arxiv.2409.02104
[9]

InSIGGRAPH Asia 2024 Conference Papers

Dynamic Gaussian Marbles for Novel View Synthesis of Casual Monocular Videos. InSIGGRAPH Asia 2024 Conference Papers. Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing

work page 2024
[10]

ACM Trans

Representing Long Volumetric Video with Temporal Gaussian Hierarchy. ACM Trans. Graph.43, 6 (2024), 171:1–171:18. Youngsik Yun, Jeongmin Bae, Hyunseung Son, Seoha Kim, Hahyun Lee, Gun Bang, and Youngjung Uh

work page 2024
[11]

InACM SIGGRAPH 2025 Conference Papers (Siggraph ’25)

Compensating Spatiotemporally Inconsistent Observations for Online Dynamic 3D Gaussian Splatting. InACM SIGGRAPH 2025 Conference Papers (Siggraph ’25). Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu

work page 2025
[12]

Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

Editable Free-Viewpoint Video Using a Layered Neural Representation.ACM Transactions on Graphics40, 4 (2021), 149:1–149:18. Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

work page 2021
[13]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

DrivingGaussian: Composite Gaussian Splatting for Surrounding Dy- namic Autonomous Driving Scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21634–21643. SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong. 10•Daheng Yin, Isaac Ding, Yili Jin, Jianxin Shi, and Jiangchuan Liu Fig

work page 2025
[14]

Our method achieves higher and more stable visual quality than baselines in most cases, demonstrating its robustness

Average visual quality (PSNR ↑ /SSIM ↑ /LPIPs ↓) over long-video sequences using our parallel pipeline with 8 GPUs (long-video experiments). Our method achieves higher and more stable visual quality than baselines in most cases, demonstrating its robustness. Lines ending prematurely for 4DGS and ST-4DGS indicate training failures due to GPU memory overflo...

work page 2025
[15]

Our method generates fewer artifacts and better preserves visual details compared to baselines, particularly in highly dynamic regions

Qualitative comparison of rendered results from the final frame of representative 9-frame clips processed in parallel using 8 GPUs (short-clip experiments). Our method generates fewer artifacts and better preserves visual details compared to baselines, particularly in highly dynamic regions. SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong

work page 2025

[1] [1]

InACM SIGGRAPH 2024 Conference Papers (SIGGRAPH ’24)

4D-Rotor Gaussian Splatting: Towards Efficient Novel View Synthesis for Dynamic Scenes. InACM SIGGRAPH 2024 Conference Papers (SIGGRAPH ’24). 1–11. Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang

work page 2024

[2] [2]

Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds

InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds. doi:10.48550/ARXIV.2403.20309 Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu

work page doi:10.48550/arxiv.2403.20309

[3] [3]

Taichi: a language for high-performance computation on spatially sparse data structures.ACM Transactions on Graphics (TOG)38, 6 (2019),

work page 2019

[4] [4]

Co- tracker3: Simpler and better point tracking by pseudo- labelling real videos

CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos. doi:10.48550/ARXIV.2410.11831 SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong. TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction•9 Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi,...

work page doi:10.48550/arxiv.2410.11831 2025

[5] [5]

InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.)

CoTracker: It Is Better to Track Together. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). 18–35. Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis

work page 2024

[6] [6]

Guillaume Le Moing, Jean Ponce, and Cordelia Schmid

3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (2023), 139:1–139:14. Guillaume Le Moing, Jean Ponce, and Cordelia Schmid

work page 2023

[7] [7]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. 2024b. ST-4DGS: Spatial-Temporally Consistent 4D Gaussian Splatting for Efficient Dynamic Scene Rendering. InACM SIGGRAPH 2024 Con...

work page 2024

[8] [8]

arXiv preprint arXiv:2409.02104 , year=

DynOMo: Online Point Tracking by Dynamic Online Monocular Gaussian Reconstruction. arXiv:2409.02104 [cs] doi:10.48550/arXiv.2409.02104 Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas

work page doi:10.48550/arxiv.2409.02104

[9] [9]

InSIGGRAPH Asia 2024 Conference Papers

Dynamic Gaussian Marbles for Novel View Synthesis of Casual Monocular Videos. InSIGGRAPH Asia 2024 Conference Papers. Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing

work page 2024

[10] [10]

ACM Trans

Representing Long Volumetric Video with Temporal Gaussian Hierarchy. ACM Trans. Graph.43, 6 (2024), 171:1–171:18. Youngsik Yun, Jeongmin Bae, Hyunseung Son, Seoha Kim, Hahyun Lee, Gun Bang, and Youngjung Uh

work page 2024

[11] [11]

InACM SIGGRAPH 2025 Conference Papers (Siggraph ’25)

Compensating Spatiotemporally Inconsistent Observations for Online Dynamic 3D Gaussian Splatting. InACM SIGGRAPH 2025 Conference Papers (Siggraph ’25). Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu

work page 2025

[12] [12]

Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

Editable Free-Viewpoint Video Using a Layered Neural Representation.ACM Transactions on Graphics40, 4 (2021), 149:1–149:18. Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

work page 2021

[13] [13]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

DrivingGaussian: Composite Gaussian Splatting for Surrounding Dy- namic Autonomous Driving Scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21634–21643. SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong. 10•Daheng Yin, Isaac Ding, Yili Jin, Jianxin Shi, and Jiangchuan Liu Fig

work page 2025

[14] [14]

Our method achieves higher and more stable visual quality than baselines in most cases, demonstrating its robustness

Average visual quality (PSNR ↑ /SSIM ↑ /LPIPs ↓) over long-video sequences using our parallel pipeline with 8 GPUs (long-video experiments). Our method achieves higher and more stable visual quality than baselines in most cases, demonstrating its robustness. Lines ending prematurely for 4DGS and ST-4DGS indicate training failures due to GPU memory overflo...

work page 2025

[15] [15]

Our method generates fewer artifacts and better preserves visual details compared to baselines, particularly in highly dynamic regions

Qualitative comparison of rendered results from the final frame of representative 9-frame clips processed in parallel using 8 GPUs (short-clip experiments). Our method generates fewer artifacts and better preserves visual details compared to baselines, particularly in highly dynamic regions. SA Conference Papers ’25, December 15–18, 2025, Hong Kong, Hong Kong

work page 2025