LiDAR + camera perception data for an AV program
An autonomous-driving team needed dense 3D cuboid and segmentation labels at a volume and consistency its vendor could not sustain.
The challenge and pain points
Perception regressions traced back to inconsistent 3D labeling and slow turnaround on edge-case scenes.
- Perception regressions kept tracing back to inconsistent 3D labeling rather than model architecture.
- The incumbent vendor could not hold consistency as volume grew, so quality drifted batch to batch.
- Edge-case scenes turned around slowly, which blocked the exact examples the perception model most needed.
- Cuboid and segmentation conventions varied between labelers, so the same object was annotated differently across frames.
- Without scenario-based gold sets, there was no objective gate to catch a bad batch before it reached training.
What Corpshore did
Stood up dedicated LiDAR pods with scenario-based gold sets and per-frame consensus review under the QA cascade.
- 1Build scenario-based gold sets
Defined gold-standard scenes for the hard cases so every batch could be measured against a fixed reference before delivery.
- 2Stand up dedicated LiDAR pods
Staffed dedicated 3D annotation pods rather than marketplace bidding, so the same trained team held one labeling convention across frames.
- 3Fix cuboid and segmentation conventions
Documented a single set of conventions for cuboids and semantic segmentation so identical objects are labeled identically across the sequence.
- 4Run per-frame consensus review
Applied per-frame consensus and expert audit under the three-tier QA cascade, catching drift before it entered the training set.
- 5Prioritize edge-case turnaround
Sequenced edge-case scenes so the model received its hardest examples on a predictable cadence instead of waiting.
The solution
Corpshore replaced variable vendor output with dedicated LiDAR pods working to one documented convention for cuboids and segmentation. Because the same trained team stays on the program, the way an object is labeled in one frame matches how it is labeled in the next, which is the consistency perception models depend on.
Scenario-based gold sets gave the program an objective gate. Every batch is measured against fixed reference scenes, and per-frame consensus review under the three-tier QA cascade catches drift before it reaches training rather than after a regression appears.
Edge-case scenes are sequenced deliberately so the perception model receives its hardest examples on a predictable cadence. The engagement delivered high-consistency 3D labels at scale, and the client reported fewer perception-layer regressions traceable to labeling.
Results
The program delivered dense 3D cuboid and segmentation labels at 97%+ accuracy against gold sets, with a labeling convention held consistent across frames and sequences. The client reported that perception-layer regressions previously traced to inconsistent labeling were reduced once the dedicated pods and scenario gold sets were in place. The exact regression figures are confidential to the client; the improvement is described qualitatively here.
| Metric | Result | Notes |
|---|---|---|
| Accuracy | 97%+ | Measured against scenario gold sets |
| Modality | LiDAR + camera | Dense 3D cuboids and segmentation |
| Label consistency | Held across frames | Single documented convention |
| Perception regressions | Reduced | Client figure confidential; qualitative |
| Edge-case turnaround | Predictable cadence | Hardest scenes sequenced deliberately |
| Staffing | Dedicated pods | Not marketplace bidding |
| Cost vs US-domestic | 50 to 70% lower | Methodology-level figure, not a client quote |
Illustrative cost index, Corpshore set to 100. The band reflects Corpshore's typical 50 to 70% cost advantage versus US-domestic providers. Representative of methodology, not exact client data.
Representative distribution of caught errors across the cascade, locking in 97%+ accuracy before delivery.
Client names and some figures are confidential. Where an exact client metric is not published, outcomes are described qualitatively and charts show representative or methodology-level data, including the three-tier QA cascade distribution and Corpshore's typical 50 to 70% cost advantage versus US-domestic providers.
Ready to scope a pilot?
Tell us your modality, volume, and languages. We'll return an indicative scope, timeline, and cost band.