Close the vulnerable-road-user safety-metric gap

Close the performance gap between S²D² and S3CNet on vulnerable-road-user intersection-over-union, particularly the gap caused by S3CNet’s higher performance on the three vulnerable-road-user classes under the single-sweep evaluation protocol.

Background

The paper introduces VRU-IoU as a metric focused on person, bicyclist, and motorcyclist classes because overall mIoU can obscure performance on rare but safety-critical road users. Although S²D² improves the refined SCPNet result, S3CNet achieves a substantially higher VRU-IoU under the single-sweep comparison. The authors explicitly identify closing this safety-performance gap as unresolved work.

References

On the metric we introduce precisely because mIoU hides the tail, a 2021 method is ahead of us, and closing that gap is open work.

— Generative Semantic Scene Completion  (2608.26737 - Chen et al., 27 Aug 2026) in Section 6, “Where the Method Loses,” paragraph “We do not lead on safety”