Checkpoint 16 · Visual QA
Classical lane detector review
A camera-grounded inspection of one-sweep and ten-sweep LiDAR lane detections. The remaining limitation is recall—not wild-line failure.
Start here
Read the review in two rows
The top row proves LiDAR presence: one sweep is depth-coloured and ten sweeps are coloured by source coverage. The bottom row contains only accepted lane curves. Clean camera imagery means the detector declined to publish a lane.
Camera projection transform: initial surveyed transform.
Comparative evidence
One sweep versus ten
Blue cells contain one contributing sweep; yellow cells contain all ten. This is a source-coverage cue, not reflectivity.
Manual visual audit
Every camera comparison, inspected
This is a visual sanity check, not an accuracy measurement: the recording has no lane ground truth.
| Camera | Scene | Visual result | Judgement |
|---|---|---|---|
| 348 | Side-looking / near-stationary | Ten-sweep window not built | Correct REJECT_REDUNDANT; the 0.03 m baseline would only amplify façade and kerb clutter. |
| 800 | Turning interval | Ten-sweep window not built | Correct REJECT_MOTION; 21.51° yaw is unsafe for this lane assumption. |
| 900 | Straight urban road | Two boundaries align with the visible road corridor in both views | Plausible positive; the left curve may represent the road edge rather than painted lane. |
| 1000 | Straight road with clear right marking | Two curves follow the visible left road boundary and right white marking | Strongest plausible positive; stable from one to ten sweeps. |
| 1200 | Straight road with visible paint | No pair published | Conservative false negative; the signal does not support both sides strongly enough. |
| 1400 | Intersection / crosswalk | No pair published | Correct negative; crossing paint is not promoted to longitudinal lanes. |
| 2600 | Straight narrow urban road | Single sweep misses; ten sweeps recover two short aligned boundaries | Clearest aggregation benefit in this sample. |
| 2800 | Straight road with asymmetric paint | No pair published | Conservative false negative; one-sided evidence is not fabricated into a pair. |
| 3000 | Intersection approach | No pair published | Correct negative. |
Three windows publish visually plausible paired boundaries, while the intersection and turn examples stay clean. Cameras 1200 and 2800 remain known misses.
Classical configuration
Selected detector settings
Validated classical defaults are preserved, with explicit geometry, aggregation, and source-plane alignment gates.
y −4–4 m
z −2–0.5 m
8° tilt limit
5 m overlap
paired around ego
Window index
Inspect every result
Filter by outcome, then open any review, LiDAR projection, accepted lane view, or equal-scale bird’s-eye comparison at full resolution.
| Start | Camera | Status | 1 sweep | 10 sweeps | Residual | Motion | Inspect |
|---|---|---|---|---|---|---|---|
| 1792 | 348 | REJECT_REDUNDANT | 0 | 0 | −2.3 ms | 0.03 m | Rejection |
| 2244 | 800 | REJECT_MOTION | 0 | 0 | −5.9 ms | 1.69 m | Rejection |
| 2344 | 900 | VALID_LANE | 2 | 2 | −2.7 ms | 3.98 m | ReviewLiDARLanesBEV |
| 2444 | 1000 | VALID_LANE | 2 | 2 | −7.8 ms | 5.46 m | ReviewLiDARLanesBEV |
| 2644 | 1200 | VALID_NO_LANE | 0 | 0 | −9.8 ms | 3.36 m | ReviewLiDARLanesBEV |
| 2844 | 1400 | VALID_NO_LANE | 0 | 0 | −3.4 ms | 4.59 m | ReviewLiDARLanesBEV |
| 4044 | 2600 | VALID_LANE | 0 | 2 | −6.7 ms | 0.53 m | ReviewLiDARLanesBEV |
| 4244 | 2800 | VALID_NO_LANE | 0 | 0 | −8.6 ms | 1.99 m | ReviewLiDARLanesBEV |
| 4444 | 3000 | VALID_NO_LANE | 0 | 0 | −2.2 ms | 3.67 m | ReviewLiDARLanesBEV |
Rejected aggregation windows
No synthetic result when the gate fails
A rejected window still shows the camera and one LiDAR sweep, but does not fabricate a ten-sweep result.
Camera views use real undecorated H.265 frames. Projection views thin points only for readability; the detector receives the full cropped cloud. There are no JSON deliverables.