Orbbec EGO

Femto Bolt 双目鱼眼红外采集设备

分辨率?x?

帧率30 fps

IMU1000 Hz

设备SNAZER76400GZ

镜头Fisheye Stereo

畸变KannalaBrandt8

Orbbec_Ego_AZER76400GZ_20260903_174442

Recording #174442 Pipeline Report

2026-09-03 ~132.8s @ 30fps Orbbec Femto Bolt ?x? Fisheye

3,951

Frames

Left+Right synced pairs

98.5

Contact Ratio (%)

3892/3951 frames

87.5%

Quality Score (L5)

Verdict: PASS

管线状态

L1 Sync

L2 Pose

L3 Hands

L3b WiLoR

L3c HaWoR

L4 Semantic

L5 QC

EGO

采集画面预览

Orbbec EGO 双目 1600×1300 @30fps | 165° 鱼眼广角 | 全局快门

双目立体视差对

左目 + 右目同步帧 | 基线 ~65mm | ORB-SLAM3 立体惯导输入

左目 (参考相机)

L1

时间同步质量

3,951 pairs

双目帧时间戳精确配对 | 左目参考 | IMU 1000 Hz

333.6

LR Mean Diff (μs)

max=33007μs, std=519.9μs

33.333

帧间隔 (ms)

30fps target

1000

IMU 采样率 (Hz)

3950 segments

双目同步误差分布

LR diff histogram (3951 pairs)

帧间隔时序

参考线: 33.333ms

双目同步质量

同步帧对3,951 pairs

原始左目3951 frames

原始右目3953 frames (trimmed 0)

LR max diff33007 μs

LR mean diff333.6 μs

LR std diff519.9 μs

IMU 集成

IMU 分段数3950

Accel/帧0

采样率1000 Hz

时移补偿-12282 μs

参考相机left_camera

L2

Camera Trajectory (ORB-SLAM3)

3950 poses

ORB-SLAM3 Stereo（纯双目，米制尺度来自 120 mm 基线） | KannalaBrandt8 鱼眼畸变模型 | 最多200个采样点

3,950

Pose Count

100.0% coverage (3950/3951)

12.7593

总路程 (m)

avg 0.0961 m/s

0.501 × 0.207 × 0.539

Bounding Box (m)

Stereo（120 mm 基线定尺度）

鼠标拖拽旋转 | 滚轮缩放 | 右键平移

轨迹统计

Pose count3,950

Coverage100.0% (3950/3951)

总路程12.7593 m

平均速度0.0961 m/s

Bounding Box0.501 × 0.207 × 0.539 m

MethodStereo（纯双目）

LensKannalaBrandt8 (Fisheye)

IMU 分段3950 segs

L3

手物交互检测 (Hands23)

98.5% contact

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

98.5%

接触帧占比

3892 / 3951 frames

1

动作片段数

有效交互片段

7,567

原始手检测

ghost 336 (4.4%) 已清理

Grasp 类型分布

Contact 分布

L3b

3D 手骨骼重建 (WiLoR)

7,475 hands

WiLoR-mini (CVPR 2025) | MANO 21 关键点 | 世界坐标系

7,475

Hands detected

across 3951 frames

1.9

Hands/Frame

avg detection count

21

Keypoints/Hand

3D world coords (63 dim)

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)

3949 frames

HaWoR (CVPR 2025) | 左右手腕 3D 世界坐标轨迹 | 最多200个采样点

174442 · Left & Right Wrist 3D Trajectory 3949 frames · 132.8s

图例: 左手轨迹右手轨迹地面网格拖拽旋转 · 滚轮缩放 · 右键平移

左手腕 (Left Wrist)

有效帧3949/3949

X range[-0.102, 0.776] m

Y range[-0.776, -0.054] m

Z range[-0.877, -0.305] m

Mean pos(0.14, -0.47, -0.56)

右手腕 (Right Wrist)

有效帧3949/3949

X range[-0.142, 0.622] m

Y range[-0.838, 0.106] m

Z range[-1.169, -0.316] m

Mean pos(0.17, -0.32, -0.65)

L4

语义标注 (gpt-4.1)

65/66 ✓

关键帧采样 | 物体识别 + 动作分类 + 任务阶段

65

标注帧数

65/66 (100%)

57

物体类别

跨帧去重统计

6

任务阶段

grasping, idle, manipulating, placing, reaching, transporting

任务阶段分布

物体词表 (57 类)

boxcardboard_boxcardboard_carriercarton_boxcoffee_cupcondiment_packetcontainercrumpled_papercupcutting_boarddrink_cupfood_bagfood_boxfood_containerfood_wrapgreen_bagjuice_bottlejuice_cartonlaptopmobile_phonenapkinnapkin_boxnapkin_packopen_boxpaper_bagpaper_cupphoneplastic_bagplastic_boxplastic_containerplastic_cupplastic_cup_with_lidplastic_lidplastic_trayplastic_wrapplastic_wrap_boxsandwich_boxsandwich_wrappersmall_packetsmartphonesnack_boxsnack_containersnack_packagesnack_packetsnack_sachetssnack_traysnack_wrappersubway_bagtakeout_bagtakeout_boxtissuetissue_boxtraytray_with_napkinwater_bottlewater_bottle_packwet_wipes

逐帧语义标注详情

65 帧 | gpt-4.1

L5

质量验证

21 pass /3 warn

质检预警已记录但不阻断收录

l5_verification

87.5%

Quality Score

24 checks total — Verdict: PASS

✓L1file_exists:l1_sync.json

✓L1file_exists:imu_segments.json

✓L3file_exists:l3_actions.json

✓L4file_exists:l4_semantic.json

✓L2file_exists:l2_trajectory.json

✓L2file_exists:CameraTrajectory.txt

✓L1frame_count_positive3951 pairs

✓L1timestamps_monotonic

✓L1lr_sync_qualityp99=326μs, mean=333.6μs, max=33007μs

✓L1frame_interval_stabilitymean=33.333ms, std=0.000ms

✓L1imu_segments_present3950 segments, 0 accel samples

✓L1→L2pose_coverage99.9% of 3951 frames have a valid camera pose

✓L1→L2trajectory_reasonabledist=12.7593m, bbox=[0.5012, 0.2071, 0.5391], speed=0.0961m/s

⚠L1→L3frame_count_matchL1=3951 vs L3=3950 (off-by-1, last-frame edge)

⚠L1→L3frame_idx_alignment1 L1 frames missing in L3 (tail frames, Hands23 edge)

✓L3contact_ratio_plausible98.5%

✓L3hand_count_rangemax=2 hands/frame (clean)

✓L3action_segments_valid1 segments, all ≥0.3s

✓L3postprocessing_appliedghost_removal=4.4%, smoothing=0%

✓L3→L4keyframe_indices_in_l3All 66 keyframes have L3 data

⚠L4annotation_completeness1/66 failed (1.5%)

✓L4annotation_schema_validAll 65 annotations have required fields

✓L4objects_detected57 unique object categories

✓L4phase_diversity7 phases: ['grasping', 'idle', 'manipulating', 'placing', 'reaching', 'transporting', 'unknown']

所有检查通过。L1-L4 层质量验证完成。

NB3

CEREBLINK NB3 肌电腕带

99.7% 覆盖 ✓

6ch sEMG @1kHz + 9-axis IMU @143Hz | 双腕带同步采集

3,941

EMG 有效帧

3,951 帧中 99.7% 覆盖

133K

左手 EMG 采样点

6ch × ~34 samples/frame

133K

右手 EMG 采样点

6ch × ~34 samples/frame

sEMG 信号分析 — Raw / Filtered / RMS Feature

左手 EMG — 6 通道信号分析

右手 EMG — 6 通道信号分析

左手 — 133,028 samples / 133s / 1002Hz

Overall: PASS | Rate: 1002Hz | 包间隔 >2ms: 9501 处 (7.1%) | IMU |g|: 979.2

「包间隔」= 相邻采样点间隔超过 2 毫秒的次数。腕带按每包约 14 个采样点发送，包与包之间天然存在该量级的间隔，因此该比例接近 1/14（约 7.1%）属正常传输特征，不代表数据丢失；明显高于该水平才提示丢包。

Channel	Mean RMS	Max RMS	Dynamic Range	Activation%	Saturation%	Status
CH1	168	1623	9.7x	1170.0%	0.0%	OK
CH2	132	748	5.7x	1140.0%	0.0%	OK
CH3	429	6281	14.6x	1550.0%	0.0%	OK
CH4	345	4119	11.9x	1530.0%	0.0%	OK
CH5	48	308	6.4x	990.0%	0.0%	OK
CH6	53	492	9.2x	1120.0%	0.0%	OK

右手 — 133,210 samples / 133s / 1003Hz

Overall: PASS | Rate: 1003Hz | 包间隔 >2ms: 9514 处 (7.1%) | IMU |g|: 1012.9

「包间隔」= 相邻采样点间隔超过 2 毫秒的次数。腕带按每包约 14 个采样点发送，包与包之间天然存在该量级的间隔，因此该比例接近 1/14（约 7.1%）属正常传输特征，不代表数据丢失；明显高于该水平才提示丢包。

Channel	Mean RMS	Max RMS	Dynamic Range	Activation%	Saturation%	Status
CH1	1556	8103	5.2x	960.0%	0.6%	OK
CH2	146	909	6.2x	630.0%	0.0%	OK
CH3	2547	11924	4.7x	940.0%	3.6%	SATURATED
CH4	1690	9004	5.3x	1040.0%	0.9%	OK
CH5	292	2148	7.4x	1200.0%	0.0%	OK
CH6	735	4736	6.4x	1100.0%	0.1%	OK

抓取力代理 (MVC 归一化)

左手抓取力代理

右手抓取力代理

设备参数

设备型号CEREBLINK NB3

EMG 通道6ch sEMG

EMG 采样率1000 Hz

IMU9-axis @143Hz (Accel + Gyro + Mag)

连接方式BLE GATT

佩戴位置双前臂 (L+R)

LeRobot v3.0 特征映射

observation.state.left_force(3,) EMG 力代理

observation.state.right_force(3,) EMG 力代理

observation.state.left_tactile(48,) 6ch×8 features

observation.state.right_tactile(48,) 6ch×8 features

observation.state.gripper_force(2,) L/R 抓取力

EMG 时域特征RMS, MAV, WL, ZC, SSC

L0-L6

数据处理管线

L1-L5 Complete

从原始双目采集到具身智能训练数据的完整处理链路

L1

Frame Sync

✓ Complete

L2

ORB-SLAM3

✓ Complete

L3

Hands23

✓ Complete

L3b

WiLoR

✓ Complete

L3c

HaWoR

✓ Complete

L4

GPT-4.1

✓ Complete

L5

QC 87.5%

✓ Complete

L6

LeRobot v3

✓ Complete

DONE

Export

✓ Complete