← Back to Blog

Faraway-Frustum: Bringing Frustum-based Point-cloud to Far-distance Object Detection

Motivation

Frustum-based 3D object detection methods — which first detect objects in 2D and then lift them to 3D using point cloud data within a frustum — have achieved strong results in near-range scenarios. However, far-distance detection remains a significant challenge:

These issues compound: a slightly misaligned 2D bounding box at 150m can produce a frustum that misses the actual 3D object entirely or includes excessive background noise.

Method: Faraway-Frustum

We extend the standard frustum-based pipeline with three key innovations tailored for far-distance objects:

1. Adaptive Frustum Expansion

Instead of using a fixed frustum size based on the 2D detection, we dynamically expand the frustum proportionally to the estimated distance. Objects predicted to be farther away receive wider frustums, ensuring the 3D points of interest are captured even with 2D detection noise.

2. Multi-scale Point Cloud Aggregation

For far-range objects, we aggregate point cloud features across multiple scales — combining the raw sparse points with voxelized representations at different resolutions. This multi-scale approach compensates for the inherent sparsity of distant LiDAR returns.

3. Distance-aware Loss Weighting

Training samples are weighted inversely by their ground-truth distance, forcing the model to pay more attention to the far-range regime where performance typically degrades.

Experimental Results

Method Near (0–50m) Mid (50–100m) Far (100m+)
Baseline Frustum-PointNet 78.4% 62.1% 34.7%
Faraway-Frustum (Ours) 78.9% 66.3% 48.2%

The proposed method achieves a +13.5% absolute improvement at far distances while maintaining near-range performance, demonstrating the effectiveness of distance-aware design in frustum-based pipelines.

Conclusion

Faraway-Frustum demonstrates that with careful architectural adaptations, frustum-based methods can be extended to handle the challenging far-distance regime critical for highway-speed autonomous driving. The distance-aware design principles introduced here are broadly applicable to other point-cloud-based detection paradigms.