scene-extrapolation-benchmark
收藏资源简介:
Scene-Extrapolation Benchmark Suite 是一个用于评估生成式新视角合成模型的综合性基准测试套件,特别关注传统逐场景3D高斯泼溅重建方法失效的挑战性场景。该套件围绕四个核心评估轴线构建:1) 外推性:评估模型在未观测视角下的生成能力,而非密集采集视角间的插值。2) 动态性:处理包含运动前景内容的场景。3) 长时程:测试模型在链式或闭环相机轨迹下,累积漂移的影响。4) 记忆/重访:评估当相机返回先前观测过的位姿时,模型是能准确回忆场景还是重新产生幻觉。套件包含12个基准测试,分布在上述轴线下,具体包括静态场景的DL3DV(50个剪辑)、作为主要动态评估的Kubric-4D(提供完美GT,无深度/尺度/形变混淆)、作为次要动态检查的NVIDIA动态场景、真实动态的SpatialVID(30个剪辑)、用于评估漂移/闭环的长时程LH-DL3DV(42个剪辑)、Stanford WorldScore聚合指标,以及分别在Kubric-4D、DL3DV和SpatialVID上评估的重访一致性(multi_visit)和返回保真度(moving_back)记忆测试。评估使用广泛的定量指标,包括全帧和动态掩码的PSNR/SSIM/LPIPS、FID、TSED、MEt3R、COLMAP位姿误差、长时程漂移指标以及专门的记忆指标。同时,基于SLURM任务记录的推理时间(秒/帧)也被纳入评估。套件以自包含、可离线可视化的形式提供,包含HTML页面、内联SVG图表、JSON数据文件和视频对比面板,用于定性和定量分析。评估涵盖了多种前沿方法,包括基于形变+修复的VACE模型、多种扩散基线、逐场景4D高斯拟合方法以及长时程世界模型,旨在提供一个统一且诚实的性能衡量标准。
The Scene-Extrapolation Benchmark Suite is a comprehensive benchmark suite for evaluating generative novel view synthesis models, with a particular focus on challenging scenarios where traditional per-scene 3D Gaussian splatting reconstruction methods fail. The suite is structured around four core evaluation axes: 1) Extrapolation: Assessing the models ability to generate views from unobserved perspectives, rather than interpolating between densely sampled viewpoints. 2) Dynamism: Handling scenes containing moving foreground content. 3) Long-term: Testing the impact of accumulated drift in chained or looped camera trajectories. 4) Memory/Revisitation: Evaluating whether the model can accurately recall a scene or re-hallucinate when the camera returns to previously observed poses. The suite includes 12 benchmark tests distributed across these axes, specifically: DL3DV for static scenes (50 clips), Kubric-4D as the primary dynamic evaluation (providing perfect ground truth without depth/scale/deformation confusion), NVIDIA dynamic scenes as a secondary dynamic check, real-world dynamic SpatialVID (30 clips), long-term LH-DL3DV for evaluating drift/loops (42 clips), the Stanford WorldScore aggregate metric, and memory tests such as multi-visit and moving-back evaluated on Kubric-4D, DL3DV, and SpatialVID. Evaluation employs a wide range of quantitative metrics, including PSNR/SSIM/LPIPS for full-frame and dynamic masks, FID, TSED, MEt3R, COLMAP pose error, long-term drift metrics, and specialized memory metrics. Additionally, inference time (seconds per frame) recorded via SLURM tasks is included. The suite is provided in a self-contained, offline-visualizable format, including HTML pages, inline SVG charts, JSON data files, and video comparison panels for qualitative and quantitative analysis. The evaluation covers various state-of-the-art methods, including deformation+inpainting-based VACE models, multiple diffusion baselines, per-scene 4D Gaussian fitting methods, and long-term world models, aiming to provide a unified and honest performance measurement standard.
数据集概述
基本信息
- 数据集名称: Scene-Extrapolation Benchmark Suite
- 许可证: CC-BY-4.0
- 标签: novel-view-synthesis, scene-extrapolation, benchmark, dynamic-nvs, long-horizon, video-diffusion
核心目标
该基准测试旨在对生成式新视角合成(NVS)在 3D高斯泼溅(3D-GS)重建失败 的四种场景下进行统一、诚实的评估:
- 外推性(Extrapolative) — 对未观测到的、保留的视角进行生成(非密集捕获之间的插值)
- 动态性(Dynamic) — 包含移动的前景内容
- 长时程(Long-horizon) — 链条式/闭环相机轨迹,会产生累积漂移
- 记忆/重访(Memory / Revisit) — 当相机返回到之前观测过的位姿时,模型是否记得场景,还是重新生成幻觉
数据集内容
可视化包包含一个自包含的离线页面(viz_pkg/index.html),分为四个部分:
- Kubric-4D 难度曲线(标题) — 无混淆因素的动态场景主要评估。内联SVG折线图,展示不同相机角度区间(30°/60°/90°/120°/180°)下的指标变化
- 覆盖矩阵(Coverage matrix) — 展示哪些方法在哪些基准上已完成评分、正在运行、不适用或待处理
- 定量结果(Quantitative) — 每个基准的指标表格,包括静态、动态、长时程和记忆/重访指标
- 定性结果(Qualitative) — 并排视频对比面板
基准测试(12个,分布在4个维度)
| 维度 | 基准 | 说明 |
|---|---|---|
| 静态 | DL3DV(50个保留片段) | 外推至未观测区域 |
| 动态 | Kubric-4D | 主要基准,无混淆因素,完美GT,按角度分档 |
| 动态 | NVIDIA Dynamic Scenes | 次要真实动态检查,12相机静态支架 |
| 动态 | SpatialVID(30个片段) | 野外动态场景 |
| 长时程 | LH-DL3DV(42个片段) | GVS风格漂移/闭环测试 |
| 长时程 | WorldScore | 斯坦福世界生成聚合 |
| 记忆/重访 | multi_visit × {Kubric-4D · DL3DV · SpatialVID} | 重访一致性 — 评估重访固定观测位姿时的GT保真度和跨访问自一致性漂移 |
| 记忆/重访 | moving_back × {Kubric-4D · DL3DV · SpatialVID} | 返回保真度 — 远离到未观测空间后返回,评估返回vs GT |
评估方法
包含多种方法,分为几类:
- 本文方法(warp+inpaint VACE):
vace14b及其微调版本 - 扩散基线:
gen3c,voyager,flexworld,viewcrafter,framecrafter,nvs_solver,viewextrapolator,trajcrafter - 逐场景3D-GS:
shape_of_motion - 长时程世界模型:
lyra2,matrix3d,spmem,vmem,matrixgame3
评估指标
- 静态: PSNR/SSIM/LPIPS(空洞/外推 + 全帧)、FID、TSED、MEt3R、COLMAP
- 动态: Kubric-4D(主要,角度分档指标)、NVIDIA动态区域掩码mPSNR/mSSIM/mLPIPS、SpatialVID
- 长时程: 漂移(ATE/旋转/闭环)、WorldScore聚合
- 记忆/重访:
mv_*(重访一致性)和mb_*(返回保真度)指标表
当前状态
这是一个实时快照(2026-06-29),部分作业仍在运行中。未评分的单元格会显示为 running / pending / N/A,从不生成虚构的数字。





