◂ Demo 列表 页内导航 概览 L1 同步 L2 轨迹 L3 交互 L3b 骨架 L3c 手腕 L4 语义 L5 质检 L6 LeRobot 管线路线图 站点 ← 返回首页 登录数据采集平台
Orbbec EGO
Femto Bolt 双目鱼眼红外采集设备
分辨率1600 x 1300
帧率30 fps
IMU1000 Hz
设备SNAZER76400GZ
镜头Fisheye Stereo
畸变KannalaBrandt8
Orbbec_Ego_AZER76400GZ_20260817_184614

Recording #184614 Pipeline Report

2026-08-17 ~70.5s @ 30fps Orbbec Femto Bolt 1600x1300 Fisheye
2,114
Frames
Left+Right synced pairs
19
Features (L6)
LeRobot v3.0 format
90.8%
Contact Ratio
1919/2114 frames
Stereo pair
左目 + 右目拼接 (Stereo Pair) 1600x650 px / fisheye
Left camera
EGO 左目 1600x1300
Right camera
EGO 右目 1600x1300
EXO iPhone
iPhone EXO (左前方) 1080x1920
L1

时间同步质量

2114 pairs

双目帧时间戳精确配对 + EXO iPhone 100% 覆盖

38.4
LR Mean Diff (us)
中间2112帧 <100us
33.333
帧间隔 (ms)
std = 0.032ms
1000
IMU 采样率 (Hz)
70455 accel samples
双目同步误差分布
LR diff histogram (2114 pairs)
帧间隔时序
参考线: 33.333ms

双目同步质量

同步帧对2114 pairs
原始左目2114 frames
原始右目2132 frames (trimmed 20)
LR max diff33,347 us
LR std diff1,025.3 us
33ms max diff 是首/末帧边界效应，中间 2112 帧均 <100us，实际同步质量优秀。

EXO 同步 (iPhone)

状态OK
Coverage2114/2114 (100%)
Offset0.64s
分辨率1920x1080
FPS29.97
CodecHEVC
EXO 总帧数2371 frames
EXO 时长79.1s
L2

Camera Trajectory (ORB-SLAM3)

1633 poses

ORB-SLAM3 Stereo-Inertial | KannalaBrandt8 鱼眼畸变模型 | 327 sampled points

1,633
Pose Count
77.2% coverage (1633/2114)
8.66
总路程 (m)
avg 0.121 m/s
0.60x0.62x0.36
Bounding Box (m)
Stereo-Inertial
鼠标拖拽旋转 | 滚轮缩放 | 右键平移

轨迹统计

Pose count1,633
Coverage77.2% (1633/2114)
总路程8.66 m
平均速度0.121 m/s
Bounding Box0.60 x 0.62 x 0.36 m
MethodStereo-Inertial
LensKannalaBrandt8 (Fisheye)
IMU Segments2113 segs, 70455 accel
L3

手物交互检测 (Hands23)

90.8% contact

Hands23 Detector (NeurIPS 2023) | 逐帧手-物体交互状态、抓取类型

90.8%
接触帧占比
1919 / 2114 frames
3
动作片段数
主段 63.7s (bimanual)
3,900
原始手检测
ghost 213 (5.5%) 已清理
Grasp 类型分布 (FPHA 6类)
Contact 分布
L3b

3D 手骨骼重建 (WiLoR)

4006 hands

WiLoR-mini (CVPR 2025) | MANO 21 关键点 | 9.9 fps

4,006
Hands detected
across 2114 frames
~1.9
Hands/Frame
avg detection count
21
Keypoints/Hand
3D world coords (63 dim)
WiLoR 手骨骼叠加可视化 — 蓝色 = 左手, 红色 = 右手, 21 MANO 关键点连线 + 腕部高亮 | 1600x1300 / h264 / 9.9fps render
骨骼图例 左手 (L) 右手 (R) 腕部关键点 输出: 21 关键点 (x,y,z) + 2D 投影坐标 / 帧
L3c

世界坐标手部轨迹 (HaWoR)

2113 frames

HaWoR (CVPR 2025) | 左右手腕 3D 世界坐标轨迹 | 106 sampled points

184614 · Left & Right Wrist 3D Trajectory 2113 frames · 70.4s
图例: 左手轨迹 右手轨迹 地面网格 (10cm) 拖拽旋转 · 滚轮缩放 · 右键平移

左手腕 (Left Wrist)

检测帧2113/2113 (100%)
X range[-2.096, 0.081] m
Y range[-0.869, 0.162] m
Z range[-1.410, -0.376] m
Mean pos(-0.38, -0.40, -0.71)

右手腕 (Right Wrist)

检测帧2113/2113 (100%)
X range[-4.619, 0.505] m
Y range[-1.005, 0.465] m
Z range[-3.348, -0.361] m
Mean pos(-0.25, -0.20, -0.79)
L4

语义标注 (GPT-4o Vision)

36/36 ✓

每 2s 采样关键帧 | 物体识别 + 动作分类 + 任务阶段

36
标注帧数
36/36 (100%)
29
物体类别
跨帧去重统计
5
任务阶段
15 种子活动标签
任务阶段分布
物体词表 (29类)
water_bottle sandwich snack_package coffee_cup container paper_bag laptop tissue_box notebook pen snack_box snack_container juice_box cup bag tape_dispenser glass_jar dispenser green_object sandwich_package snack_wrappers snack_packet paper_cup backpack paper box phone drink_cup +1 more
逐帧语义标注详情
36 帧 | GPT-4o Vision | 每 2s 采样
L5

质量验证

23 pass / 1 fail
95.8%
Quality Score
24 checks total — 1 failure (L1 boundary effect, known)
✓L1file_exists:l1_sync.json
✓L1file_exists:imu_segments.json
✓L2file_exists:l2_trajectory.json
✓L2file_exists:CameraTrajectory.txt
✓L3file_exists:l3_actions.json
✓L4file_exists:l4_semantic.json
✓L1frame_count_positive2114 pairs
✓L1timestamps_monotonic
✗L1lr_sync_qualitymax=33347us > 5ms limit
✓L1frame_interval_stabilitymean=33.333ms, std=0.032ms
✓L1imu_segments_present2113 segs, 70455 accel
✓L1→L2pose_coverage1633/2114 (77.2%)
✓L1→L2trajectory_reasonable8.66m, 0.12m/s
✓L1→L3frame_count_match2114 frames
✓L1→L3frame_idx_alignment2114 indices aligned
✓L3contact_ratio_plausible90.8%
✓L3hand_count_rangemax=2 hands/frame (clean)
✓L3action_segments_valid3 segments, all ≥0.3s
✓L3postprocessing_appliedghost=5.5%, smooth=0%
✓L3→L4keyframe_indices_in_l336 keyframes have L3 data
✓L4annotation_completeness36/36 (100%)
✓L4annotation_schema_validAll 36 annotations have required fields
✓L4objects_detected29 unique object categories
✓L4phase_diversity5 phases: holding, idle, manipulating, placing, reaching
lr_sync_quality 的 FAIL 是边界效应，实际同步质量优秀。L1-L4 层全部通过。
L6

LeRobot v3.0 Export

19 features
v3.0
格式版本
LeRobot (HuggingFace)
2,114
总帧数
1 episode · 30fps
19
特征维度
3 video + 8 state + 4 action + 4 meta

Dataset Info

FormatLeRobot v3.0
repo_idbasematrix/ego-binocular-v1
robot_typeorbbec_ego_binocular
Episodes1
Frames2,114
Tasks1
FPS30
Chunk size1000

Video Channels (3)

ego_left
1600x1300 x 3ch
ego_right
1600x1300 x 3ch
exo_iphone
1920x1080 x 3ch
三视角视频预览
ego_left
video ego_left
ego_right
video ego_right
exo_iphone
video exo_iphone

Feature Table (19 features)

FeatureTypeShapeCategoryDescription
timestampfloat32[1]metaFrame timestamp (sec)
frame_indexint64[1]metaSequential frame index
episode_indexint64[1]metaEpisode identifier
indexint64[1]metaGlobal sample index
task_indexint64[1]metaTask identifier
observation.images.ego_leftvideo[3, 1300, 1600]videoLeft fisheye camera
observation.images.ego_rightvideo[3, 1300, 1600]videoRight fisheye camera
observation.images.exo_iphonevideo[3, 1080, 1920]videoiPhone EXO camera
observation.state.camera_posefloat32[7]state6DoF pose (quat+xyz)
observation.state.left_wristfloat32[3]stateLeft wrist 3D position
observation.state.right_wristfloat32[3]stateRight wrist 3D position
observation.state.left_keypoints_3dfloat32[63]state21 MANO keypoints (x,y,z)
observation.state.right_keypoints_3dfloat32[63]state21 MANO keypoints (x,y,z)
observation.state.left_contactint64[1]stateLeft hand contact flag
observation.state.right_contactint64[1]stateRight hand contact flag
action.left_wrist_deltafloat32[3]actionLeft wrist delta (m)
action.right_wrist_deltafloat32[3]actionRight wrist delta (m)
action.left_graspint64[1]actionLeft hand grasp type
action.right_graspint64[1]actionRight hand grasp type
L0-L6

数据处理管线

All Complete

从原始双目采集到具身智能训练数据的完整处理链路 — 全部绿灯

EXO

iPhone EXO 视频

左前方视角
iPhone EXO 视频 (左前方) — 1080x1920 / HEVC / 29.97fps / 15s clip — 与 EGO 帧精确同步 (offset 0.64s)