KTH/SynFlow
收藏资源简介:
SynFlow数据集是一个用于LiDAR场景流估计的合成基准数据集,通过CARLA模拟器生成,旨在支持自动驾驶研究。数据集提供密集的3D场景流真实标签,其中自车在多个CARLA城镇中沿预定义路线行驶,LiDAR以10Hz频率捕获点云。每个场景存储为单独的HDF5文件,持续时间约为20秒(64束LiDAR约200帧)或30秒(32束LiDAR约302帧)。动态对象(如车辆、行人)具有实例级刚性流标签,静态背景点则通过自运动补偿流处理。数据集包括四种配置:1k(约271,148帧,覆盖城镇06、07、10的动脉道路和复杂交叉口)、2k(约449,407帧,覆盖城镇01-05的环岛和多车道交叉口)、3k(1k和2k的组合,约720,555帧)和4k(2k和城镇12的组合,约939,083帧)。数据格式与OpenSceneFlow兼容,包含点云、姿态矩阵、场景流向量、实例ID、类别索引和地面掩码等信息。数据集可用于训练和评估场景流估计模型,并提供了预训练模型检查点。
The SynFlow dataset is a synthetic LiDAR scene flow benchmark collected using the CARLA simulator, designed for autonomous-driving research. It provides dense 3D scene-flow ground truth, where an ego vehicle drives along pre-defined routes in multiple CARLA towns while a LiDAR captures point clouds at 10 Hz. Each scene is stored as a separate HDF5 file and spans roughly 20 seconds (~200 frames for 64-beam LiDAR) or 30 seconds (~302 frames for 32-beam LiDAR). Dynamic objects (e.g., vehicles, pedestrians) receive instance-level rigid-body flow labels, while static background points are compensated with ego-motion flow. The dataset includes four configurations: 1k (approximately 271,148 frames, covering Town06, 07, 10 with arterials and complex junctions), 2k (approximately 449,407 frames, covering Town01–05 with roundabouts and multi-lane intersections), 3k (combination of 1k and 2k, approximately 720,555 frames), and 4k (combination of 2k and Town12, approximately 939,083 frames). The data format is compatible with OpenSceneFlow and includes point clouds, pose matrices, scene flow vectors, instance IDs, category indices, and ground masks. It supports training and evaluation of scene flow estimation models and comes with pretrained model checkpoints.
数据集概述
SynFlow 是一个用于自动驾驶研究的合成 LiDAR 场景流(scene flow)基准数据集,由 CARLA 模拟器生成,格式与 OpenSceneFlow 兼容。
基本信息
- 任务: 其他(场景流估计)
- 模态: 3D
- 语言: 英语
- 许可证: CC-BY-4.0(SynFlow-4k 数据集及基于其训练的模型检查点)
- 标签: scene-flow, lidar, autonomous-driving, carla, synthetic, 3d
数据集内容
- 自车车辆在多个 CARLA 城镇中沿预定义路线行驶,LiDAR 以 10 Hz 频率采集点云。
- 每个场景存储为一个独立的 HDF5 文件,时长约 20 秒(64 线 LiDAR 约 200 帧)或约 30 秒(32 线 LiDAR 约 302 帧)。
- 动态物体(车辆、行人、骑行者等)获得实例级刚体流标签;静态背景点获得自车运动补偿流。
数据划分
| 划分 | 标注帧数 | 文件夹名称 | 组成 | 存储大小 (GB) |
|---|---|---|---|---|
| 1k | 271,148 | town-06-07-10 |
Town06, 07, 10 | 214G |
| 2k | 449,407 | town-01-05 |
Town01–05 | 419G |
| 3k | 720,555 | town-06-07-10 + town-01-05 |
Town06-07-10, Town01-05 | 633G |
| 4k | 939,083 | town-01-05 + town-12 |
Town01-05, Town12 | 986G |
文件结构
- 文件命名:
scene-{town_id}{channels}{route_id:04d}{scene_split:02d}.h5town_id:CARLA 城镇编号(如01,12)channels:LiDAR 线束(32,64)route_id:4 位路线索引scene_split:2 位长路线划分编号
- 提供索引文件
index_total.pkl,包含所有帧的[scene_id, timestamp]对。 - 提供基于 DeltaFlow 骨干网络的预训练模型检查点(
synflow-4k-longadp.ckpt和synflow-real-longadp.ckpt)。
数据采集
- 模拟器: CARLA 0.9.16
- 传感器:
sensor.lidar.ray_cast_semantic - 坐标系统: 右手系(RHS)
- 自车: Tesla Model 3
- NPC 数量: 每场景约 70 辆车辆 + 约 80 名行人
- 城镇: Town01–Town10, Town12
HDF5 模式
每个 HDF5 文件包含一个场景,每帧数据存储在一个以微秒时间戳命名的组(group)下。每个组包含:
| 键 | 形状 | 数据类型 | 描述 |
|---|---|---|---|
lidar |
(N, 3) | float32 | 传感器坐标系下的点云 (X, Y, Z) |
pose |
(4, 4) | float64 | 自车 4×4 变换矩阵(世界 ← 自车) |
flow |
(N, 3) | float32 | 从时刻 t 到 t+1 的场景流向量 |
flow_is_valid |
(N,) | bool | 每个点的流有效性掩码 |
flow_category_indices |
(N,) | uint8 | 语义类别索引(0 表示背景/静态) |
flow_instance_id |
(N,) | int16 | 实例 ID(-1 表示背景,正数表示动态物体) |
ground_mask |
(N,) | bool | 地面点掩码(True 为地面) |
引用
bibtex @article{zhang2026synflow, author = {Zhang, Qingwen and Zhu, Xiaomeng and Jiang, Chenhan and Jensfelt, Patric}, title = {{SynFlow}: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data}, journal = {arXiv preprint arXiv:2604.09411}, year = {2026}, }




