Inspect Multimodal Perception

Deterministic inspection for synchronized 6-camera surround vision and 3D LiDAR point clouds.

02 · INTERACTIVE PRODUCT PROOF/CURATED RUNTIME RECORD

Four verification surfaces. Zero hidden assumptions.

run-20260819-0061-0012 · Scene #0061 · 3D Viewer
SAMPLE: scene-0061
WORLD/FLU|POINTS: 34,812BOXES: 8 PREDICTED
EGO_VEHICLE
Car · 0.89 · 13.1m
Truck · 0.94 · 24.5m
Pedestrian · 0.78 · 8.2m
LiDAR Scan: 32-Beam
6× Camera Frustums
10.0 Hz SYNC

Interactive 3D LiDAR point cloud inspection, oriented bounding box verification, and camera frustum alignment with millisecond playback scrubbing.

● 03 · SYSTEM ARCHITECTURE/VERIFIED API PAYLOAD

Deterministic Contract. Immutable Results.

Every inference execution returns an immutable standard response format. No hidden parameters, no client-side heuristics.

GET /api/v1/runs/{id}
200 OK
{
  "prediction": {
    "run_id": "run-20260819-0061-0012",
    "frame_index": 12,
    "sample_id": "ca9a8385f103460fb15ff52c35e93a89",
    "checkpoint_id": "bevfusion_lidar_cam_nuscenes",
    "image_width": 1600,
    "image_height": 900,
    "object_count": 8,
    "inference_latency_ms": 142.6,
    "preprocess_ms": 18.2,
    "infer_ms": 112.4,
    "postprocess_ms": 12.0,
    "e2e_ms": 142.6,
    "boxes": [
      {
        "object_id": "obj-001",
        "class_name": "car",
        "score": 0.89,
        "center_xyz": [4.12, 12.45, -0.32],
        "size_lwh": [4.82, 1.95, 1.62],
        "yaw": 0.35,
        "distance_m": 13.11,
        "image_polygon": [
          { "x": 642.0, "y": 420.0 },
          { "x": 810.0, "y": 512.0 }
        ]
      }
    ]
  },
  "camera_name": "CAM_FRONT",
  "review_status": "GENERATED",
  "review_history": []
}

Immutable Artifacts

Inference predictions are cryptographically bound to sample IDs and execution timestamps. Results cannot be mutated retroactively.

Coordinate Rigor

Oriented bounding box coordinates (`center_xyz`, `size_lwh`, `yaw`) conform strictly to vehicle FLU conventions and real camera calibration matrices.

Explicit Metric Scope

Every evaluation metric is badged with its protocol subset (e.g. `nuScenes val-81`), eliminating unverified generalized claims.

● 04 · WORKFLOW PIPELINE/END-TO-END DATAFLOW

From Scene to Review

A continuous, deterministic processing loop transforming raw surround sensor streams into validated 3D perception artifacts.

STAGE 1A
6× Camera Ring
1600×900 · 12 FPS
STAGE 1B
32-Beam LiDAR
35k pts · 10 Hz
STAGE 2
Dual-Branch Fusion
BEVFusion Backbone
STAGE 3
3D Detection Head
Oriented BBoxes
STAGE 4
Rerun Viewer
WebAssembly 3D
STAGE 5
Human Review
Immutable Audit
● 05 · PRODUCT SURFACES/SANDBOX WORKSPACES

Everything Connected in One Run

Five specialized interfaces sharing one deterministic run identity, eliminating disconnected spreadsheets and fragmented viewer tools.

RUNS REPOSITORY81 SCENES
run-20260819-0061-0012READY
run-20260819-0061-0011READY
run-20260819-0061-0010REVIEWING

Runs Workspace & Execution Dashboard

Full frame inventory, score threshold controls, asynchronous run dispatching, and real-time execution telemetry in a unified cockpit.

RERUN CANVAS
• 32-beam pointcloud
• ego_pose FLU synced
• 6 frustums mapped

Rerun 3D WebViewer

Instant WebGL/WebGPU 3D playback scrubbing, point cloud slicing, and oriented 3D bounding box validation directly in-browser.

EVAL SUBSET: val-81

0.702 mAP / 0.729 NDS Benchmark Matrix

Official 10-class nuScenes detection metrics calculated strictly across the val-81 benchmark subset with center-distance matching.

6× PINHOLE PROJECTION

6× Surround Camera & BEV Map

Privacy-safe surround video synchronization, multi-camera pinhole projections, and top-down BEV vehicle-centric verification.

GENERATED → REVIEWED

Review Queue & Audit Trail

Enforced separation between model engineering and reviewer sign-off, recording immutable audit records for safety compliance.

● 06 · EVALUATION & EVIDENCE/VERIFICATION PAIRS

Semantic Ground Truth vs Prediction

Audit individual detections against ground truth annotations with center-distance validation and classification status.

EVAL SUBSET: nuScenes val-81
PAIR #01
MATCHED
PREDICTION
car · 89% · 13.11m
GROUND TRUTH
car · 13.05m
PAIR #02
MATCHED
PREDICTION
truck · 94% · 24.52m
GROUND TRUTH
truck · 24.48m
PAIR #03
MATCHED
PREDICTION
pedestrian · 78% · 8.24m
GROUND TRUTH
pedestrian · 8.19m
PAIR #04
FALSE POSITIVE
PREDICTION
traffic_cone · 42% · 18.9m
GROUND TRUTH
None
PAIR #05
MISS
PREDICTION
None
GROUND TRUTH
barrier · 32.1m

Open the Sandbox

Inspect deterministic 6-camera surround vision and 3D LiDAR point clouds, verify spatial detections, or audit safety review records.