EvTrajGS: Accurate and Efficient 3D Gaussian Splatting from Unposed Event Streams

Traditional 3D reconstruction from event cameras often faces a trade-off between efficiency and accuracy. Methods that rely on fixed, off-the-shelf pose estimates are fast but prone to low fidelity due to cumulative errors. Conversely, SLAM-style methods that perform joint pose-scene optimization achieve higher quality but are computationally expensive due to repeated tracking and dense bundle adjustment. This paper asks: can we achieve high-fidelity 3D Gaussian Splatting from unposed event streams without the overhead of traditional SLAM pipelines?

EvTrajGS introduces a continuous-time trajectory representation for camera motion. Instead of treating camera poses as isolated, independent variables, the authors use a Continuous Time-to-Pose Function (CTPF) to parameterize motion as a smooth, continuous function. This allows the system to query camera poses at any timestamp.

To ensure temporal consistency during optimization, the authors propose Temporally Coupled Pose Integration (TCPI). This technique aggregates neighboring trajectory states within a local temporal window into a single, coupled pose. By capturing local motion trends, TCPI guides pose updates more reliably than independent optimization. Additionally, the authors implement a loss-reweighted event sampling strategy that focuses computational resources on under-reconstructed intervals, further improving the quality of the final 3D Gaussian Splatting model.

EvTrajGS significantly outperforms existing methods in both geometric reconstruction quality and pose estimation accuracy. On synthetic and real-world datasets, the model achieved a 3.8 dB improvement in PSNR, a 0.1 increase in SSIM, and a reduction in ATE RMSE of over 40%. Notably, the framework is approximately 7 times faster than current state-of-the-art SLAM-style approaches, demonstrating that it effectively breaks the accuracy-efficiency trade-off.