scene-db

The problem

You have terabytes of rosbag / dataset logs. Somewhere in there are the 30-second clips where your system failed — but finding them means scrubbing through hours of data.

Localization drift after GPS outage

Your EKF diverged under a tunnel, but the log is 2 hours long. Which 10 seconds matter?

Missed detection at dusk

Perception worked in daylight but silently dropped detections as lighting changed. Where exactly?

Sudden stop = which scene?

The vehicle braked hard. Was it a false positive from the planner, or a real obstacle the detector caught late?

"It works on my data"

You need to share the exact failure scene with a teammate — not a 50GB bag file.

Edge cases you can surface

scene-db chunks your logs, computes motion features, and generates searchable captions. Here are real queries your team can run:

All Localization Perception

Localization

High yaw rate — IMU heading drift

Sharp turns stress the EKF heading estimate. IMU gyro bias and wheel slip both accumulate here. Found 30+ deg/s peaks in KITTI and PPC data.

$ scene-db search --min-yaw 20 --sort yaw

Localization

High speed — GPS latency amplified

At 77 km/h, a 50ms GPS delay = 1.1m position error. LiDAR scan distortion from ego-motion is also maximal.

$ scene-db search --min-speed 60 --sort speed

Localization

Low-speed maneuvering — dead-reckoning drift

Parking, U-turns, crawling in traffic. Wheel odometry resolution limits dominate. GPS multipath in urban canyons.

$ scene-db search --max-speed 5 --sort speed

Localization

Start from stop — GNSS reacquisition

After standstill, initial velocity is noisy. GNSS often reacquires with a position jump. IMU integration restarts.

$ scene-db search "stationary"

Localization

Loop closure — place revisit

Does the trajectory return to its starting point? scene-db detects loops and counts revisits for SLAM evaluation.

$ scene-db sequences

Perception

Hard braking — pitch shift

Braking tilts the vehicle forward. LiDAR FOV shifts, camera horizon drops. Detected 2.5+ m/s² events in KITTI.

$ scene-db search --min-decel 2.0 --sort decel

Perception

Approach to stop — tracking handoff

Transitioning from motion-based to static perception. Object trackers often drop detections in this zone.

$ scene-db search "slowly" --min-decel 1.0

Perception

Scenes with poor visibility

Low contrast, overexposure, rain on lens. VLM captions catch what rule-based features miss.

$ scene-db search -s "dark road with glare"

Perception

Crowded intersections

Multiple pedestrians, cyclists, turning vehicles — the combinatorial explosion that detectors struggle with.

$ scene-db search -s "busy intersection pedestrians"

Loc + Per

Yaw + brake combo — multi-axis stress

Simultaneous turning and braking. All sensor modalities stressed: IMU coupling, wheel slip, pitch + yaw change.

$ scene-db search --min-yaw 10 --min-decel 1.0

Loc + Per

Tunnels & underpasses

GPS denied + lighting transition. Localization falls back to LiDAR/IMU while cameras adapt to darkness.

$ scene-db search -s "tunnel entrance dark"

Loc + Per

Automatic detection

Don't know what to search for? Let scene-db find edge cases automatically with rule-based heuristics.

$ scene-db edge-cases --severity critical

Real edge cases from KITTI

These scenes were found by running scene-db on 2912 scenes across 9 datasets (KITTI, nuScenes, GLIM, Cartographer, PPC, AIST Park, Flatwall, AlienGo). Each GIF is a 5-second chunk extracted by the tool.

All Localization Perception Loc + Per

Intersection turn — localization stress test

drive_0009 frame 196-244. High yaw rate with moderate speed. EKF heading estimate is under maximum stress. IMU bias and wheel slip both contribute to drift here.

Localization yaw 25.4 °/s 15 km/h

Continuous curve — dead-reckoning drift

drive_0005 frame 0-48. Low speed + sustained yaw rate. Wheel odometry accumulates lateral error. GNSS multipath likely in this residential area with trees.

Localization yaw 10.9 °/s 13 km/h

Approach to stop — perception critical zone

drive_0011 frame 98-146. Vehicle decelerating to stop. This is where late detection of obstacles or traffic signals has the highest consequence.

Perception decel 1.7 m/s² 8 km/h

Sharp turn + braking — worst-case localization

drive_0014 frame 0-48. Peak yaw rate in the entire dataset. Simultaneous turning and braking maximizes IMU error coupling. Wheel odometry is unreliable here.

Localization yaw 30.0 °/s decel 1.9 m/s² 23 km/h

High-speed braking — 58 km/h deceleration

drive_0015 frame 0-48. Fastest sequence in the dataset. At 16 m/s, GPS update latency and LiDAR scan distortion from ego-motion become significant.

Loc + Per decel 1.8 m/s² 58 km/h

Coming to stop — perception handoff zone

drive_0009 frame 392-440. Vehicle decelerating to a full stop. The transition from motion-based to static perception is where tracking often drops objects.

Perception decel 1.1 m/s² 1 km/h

Highway braking — sensor fusion latency

drive_0001 frame 0-48. Higher speed with braking events. At 43 km/h, even 100ms of fusion latency means 1.2m of position uncertainty.

Loc + Per decel 1.1 m/s² 43 km/h

Start from stop — static-to-dynamic transition

drive_0018 frame 147-195. After 15 seconds stationary, the vehicle begins to move. Initial velocity estimate is noisy. GNSS often reacquires with a jump here.

Localization yaw 1.4 °/s 1 km/h

77 km/h — fastest in dataset, GPS latency critical

drive_0027 frame 147-187. At 21 m/s, a 50ms GPS update delay = 1.1m position error. LiDAR motion distortion is maximal. Any scan matching has to compensate aggressively.

Localization 77 km/h

Hardest braking — 2.5 m/s² from 43 km/h

drive_0019 frame 147-195. Strongest deceleration in the dataset. Pitch angle changes during braking shift LiDAR FOV and camera horizon.

Loc + Per decel 2.5 m/s² 43 km/h

Yaw + brake combo — worst-case multi-axis stress

drive_0019 frame 392-440. Simultaneous high yaw rate and braking at near-stop speed. All sensor modalities are stressed: IMU coupling, wheel slip, and low-speed GPS noise.

Loc + Per yaw 13.2 °/s decel 2.2 m/s² 2 km/h

Sharp turn #2 — low-speed yaw peak

drive_0046 frame 98-124. Near-maximum yaw rate at very low speed. Wheel encoder resolution limits and IMU gyro bias dominate at this speed.

Localization yaw 29.4 °/s 15 km/h

Intersection turn — multi-lane crossing

drive_0029 frame 294-342. Turning through a complex intersection. Map matching ambiguity is highest here — multiple lane hypotheses.

Localization yaw 27.8 °/s 22 km/h

Hard braking — first KITTI hard-brake detected

drive_0061 frame 637-685. Only hard braking event found across 25 KITTI sequences. Pitch change shifts LiDAR FOV and camera horizon simultaneously.

Loc + Per hard braking 21 km/h

Workflow: find → search → export → fix

Ingest once, query forever. Export just the frames you need for debugging or retraining.

scene-db

# 1. Ingest a KITTI sequence (108 frames → 3 chunks)
$ scene-db ingest /data/kitti/2011_09_26_drive_0001_sync
Ingesting ... Done. Created 3 scene chunks.

# 2. Search for low-speed scenes (localization edge case)
$ scene-db search "moving slowly"
Found 1 scene(s):

  [kitti_..._sync_002]
    vehicle moving slowly, 19 km/h, traveled 4.9 m
    frames 98-107

# 3. Semantic search (VLM + embedding)
$ scene-db search -s "vehicle decelerating near parked cars"
Found 2 scene(s):

  [kitti_..._sync_002] (score: 0.847)
    residential street, parked cars on both sides, vehicle slowing down

# 4. Export the scene for debugging
$ scene-db export --id kitti_..._sync_002 -o ./debug_scene
Exported 50 files to ./debug_scene

Built for robotics engineers

No infra required

SQLite on disk. No Docker, no Postgres, no Elasticsearch. Install with pip, run from your terminal.

VLM captioning

Optional GPT-4o integration generates rich scene descriptions from camera images. Falls back to rule-based if no API key.

Semantic search

Embedding-based similarity search with sentence-transformers. Find scenes by meaning, not just keywords.

Export for replay

Extract the exact images, point clouds, and IMU data for a scene. Feed it straight into your replay pipeline.

Multi-dataset

KITTI and nuScenes today. Same SceneChunk model regardless of source. Easy to add your own format.

Four commands

ingest index search export — no YAML configs, no pipeline DSLs, no ceremony.

Which data for your LiDAR SLAM?

Basic sanity check

GLIM os1_128 (491 MB, 115s)
Ouster OS1-128 + IMU. Small, fast to download. Confirm your pipeline runs.

Aggressive dynamics

AIST Park (2.1 GB, 144s)
Max decel 11.2 m/s² across all datasets. Repeated hard braking events (8.4, 7.2, 6.9 m/s²). Tests if your ESKF/EKF tracks through violent acceleration changes. Ouster + IMU.

LiDAR degeneration

Flatwall (306 MB, 33s)
Wall-only environment where LiDAR scan matching degenerates. Without IMU, localization fails. The critical test for IMU-LiDAR coupling.

Long-term drift

Cartographer 3D (9.3 GB, 20min)
20 minutes of continuous operation. IMU-only speed drifts to 6000+ km/h. Your SLAM must prevent this.

Loop closure

PPC Tokyo run1/run2
9.9 km with loop detected (2m closure). 1386 revisits. RTK-GNSS ground truth.

High yaw rate

KITTI drive_0014 / PPC Tokyo run3
Up to 30.2 deg/s. Intersection turns that stress heading estimation.

Quadruped walking — extreme IMU

AlienGo (774 MB, 344s)
Four-legged robot with Livox + T265 camera + IMU. Walking gait creates decel 29,693 m/s² and yaw 45,118 deg/s — 1000x more than any vehicle. The ultimate stress test for IMU preintegration and LiDAR-Visual-Inertial fusion.

scene-dbFind the failures hiding in your driving logs

The problem

Localization drift after GPS outage

Missed detection at dusk

Sudden stop = which scene?

"It works on my data"

Edge cases you can surface

High yaw rate — IMU heading drift

High speed — GPS latency amplified

Low-speed maneuvering — dead-reckoning drift

Start from stop — GNSS reacquisition

Loop closure — place revisit

Hard braking — pitch shift

Approach to stop — tracking handoff

Scenes with poor visibility

Crowded intersections

Yaw + brake combo — multi-axis stress

Tunnels & underpasses

Automatic detection

Real edge cases from KITTI

Intersection turn — localization stress test

Continuous curve — dead-reckoning drift

Approach to stop — perception critical zone

Sharp turn + braking — worst-case localization

High-speed braking — 58 km/h deceleration

Coming to stop — perception handoff zone

Highway braking — sensor fusion latency

Start from stop — static-to-dynamic transition

77 km/h — fastest in dataset, GPS latency critical

Hardest braking — 2.5 m/s² from 43 km/h

Yaw + brake combo — worst-case multi-axis stress

Sharp turn #2 — low-speed yaw peak

Intersection turn — multi-lane crossing

Hard braking — first KITTI hard-brake detected

Workflow: find → search → export → fix

How it works

Raw Logs

Chunk

Extract

Caption

Index

Search

Built for robotics engineers

No infra required

VLM captioning

Semantic search

Export for replay

Multi-dataset

Four commands

9 datasets, 2912 scenes, 168K frames

Which data for your LiDAR SLAM?

Basic sanity check

Aggressive dynamics

LiDAR degeneration

Long-term drift

Loop closure

High yaw rate

Quadruped walking — extreme IMU

scene-db
Find the failures hiding in your driving logs