Shared novel-view trajectory
Left: Raw 3DGS · Right: blur-aware 3DGS. Both render the same sharp novel-view camera at every frame; this is not a slideshow of test images.
Motion blur：Raw 3DGS vs typed blur route
同一 Ujikintoki 场景、同一 30 个训练位姿与 5 个 clean held-out views；唯一变化是训练图像中的 mild/strong motion blur，以及是否使用全视图 camera-trajectory blur operator。
先看结论
当前实现没有超过 Raw 3DGS：mild 为 −0.52 dB，strong 为 −1.26 dB。两个强度上 PSNR、SSIM 和 LPIPS 都同向变差，因此这不是单一 metric 的偶然波动。下面的视频和逐视图结果用于定位是否来自过度平滑、错误的 blur trajectory，或训练/测试 image-formation mismatch。
mild protocol · 30 train / 5 clean test
Mild motion blur
Blur-aware loses 0.52 dB
MethodPSNR ↑SSIM ↑LPIPS ↓
Raw 3DGS24.3230.7720.387
Blur-aware23.8030.7410.461
Blur − raw-0.520-0.030+0.074
Five clean held-out views
Held-out view 1 blur − raw -2.18 dB
Held-out view 2 blur − raw -1.06 dB
Held-out view 3 blur − raw +0.56 dB
Held-out view 4 blur − raw +0.46 dB
Held-out view 5 blur − raw -0.38 dB
strong protocol · 30 train / 5 clean test
Strong motion blur
Blur-aware loses 1.26 dB
MethodPSNR ↑SSIM ↑LPIPS ↓
Raw 3DGS21.9390.7350.464
Blur-aware20.6840.6990.538
Blur − raw-1.256-0.035+0.073
Shared novel-view trajectory
Left: Raw 3DGS · Right: blur-aware 3DGS. Both render the same sharp novel-view camera at every frame; this is not a slideshow of test images.
Five clean held-out views
Held-out view 1 blur − raw -1.76 dB
Held-out view 2 blur − raw -1.51 dB
Held-out view 3 blur − raw -0.93 dB
Held-out view 4 blur − raw -0.91 dB
Held-out view 5 blur − raw -1.18 dB
共同轨迹
轨迹由排序后的 30 个训练相机做平滑位姿插值并 ping-pong 闭环；mild/strong、Raw/blur-aware 全部使用完全相同的 novel-view poses。
