Self-driving cars see the world through LiDAR, a laser scanner that returns a cloud of 3D points. Nearby objects get many points and distant ones get few. Most detectors chop this cloud into a regular grid of small cubes called voxels. That forces one compromise cell size on everything: too coarse to resolve a pedestrian at 10 m, and wastefully fine for the empty space at 60 m. A recent fix is to stretch space before gridding. Nearby points are pushed apart so a standard grid effectively becomes finer up close, and the detector itself does not change. This gives very large gains on small objects. What nobody has asked is whether the stretch can be pushed harder than current methods dare, and whether the limit on that is really an optimization problem. A network that starts from random weights is shown a heavily distorted scene on its very first step.
We propose to introduce the stretch gradually during training. Training starts on an undistorted grid, and the warp strength is raised step by step to a target stronger than the best fixed setting. We then test, under tightly controlled and repeated runs, whether this reaches small-object accuracy that no fixed warp reaches at the same memory cost. The two-sentence version is simple. Distance-adaptive voxelization is currently a fixed preprocessing choice. We treat it as a training schedule, and we predict that doing so moves the best achievable accuracy–memory trade-off, not just the speed of convergence.
Core Claim
Radial Scaling Voxelization for Accurate Small Object 3D Detection (RSV) applies a radial warp $s(r)$ to point coordinates before voxelization. The warp expands space near the sensor and keeps the grid Cartesian, so SECOND, PointPillars and Voxel-RCNN run on the warped space unchanged. Boxes are mapped back to metric space with a Newton–Raphson solver. RSV's warp parameters ($s_{max}$, $\beta$, $r_{far}$) are fixed from the first training step.
Our claim is that fixed-warp training has an optimization ceiling. As $s_{max}$ grows, a fixed-warp detector's accuracy should first rise and then fall, even though stronger warps give strictly more near-range resolution per byte. A schedule that anneals $s_{max}$ from 1 (no warp) to a target should keep improving past the fixed-warp peak.
The motivating evidence comes from implicit 3D shape learning. Curriculum DeepSDF found that asking an untrained network for fine geometry from the start blurs thin structures. Moving from coarse to fine error tolerance cut mean Chamfer Distance by 32.3% relative. That work concerns single ShapeNet objects represented as continuous functions, not voxel grids of outdoor scenes. We carry over its method (coarse-to-fine annealing of a spatial fidelity parameter) and test whether its finding holds in a regime where the "fine" end costs memory.
The curriculum theory in A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula constrains the design. That paper shows easy-to-hard training beats uniform mixing only with a feasible starting point, moderate jumps between stages, and enough samples per stage. Its setting is iterative self-improvement of language models, so its conditions are a hypothesis for us, not a guarantee. They do make a sharp prediction. A single jump from unwarped to full warp should underperform gradual annealing, and very fine-grained annealing should give no extra benefit.
Research Program
We will build one warp module with a single interface: it takes a point cloud and a schedule value $\alpha \in [0,1]$ and returns warped coordinates and the inverse map. The warp strength follows
$$s_{max}(t) = 1 + \alpha(t)\,(s^{*}_{max} - 1),$$
where $t$ is the training step, $\alpha(t)$ rises from 0 to 1 under a chosen schedule, and $s^{*}_{max}$ is the target strength.
Because the warp changes during training, the regression targets must change with it. Ground-truth boxes are re-warped with the current $s_{max}(t)$ at every step. Evaluation always uses the final warp and the RSV inverse solver.
Modularity plays a practical role here, not a showcase role. The interface pattern follows the backbone/loss factories of FaceX-Zoo: A PyTorch Toolbox for Face Recognition. Keeping the schedule, warp and detector separate lets the identical schedule plug into three detectors and, as a secondary test, into the radial interval spacing of NUC-Net: Non-Uniform Cylindrical Partition Network for Efficient LiDAR Semantic Segmentation. NUC-Net places fine radial intervals near the sensor and coarse ones far away on a fixed arithmetic progression. We would anneal that progression from uniform toward its fixed setting.
One further ingredient borrows from How does Multi-Task Training Affect Transformer In-Context Capabilities? Investigations with Function Classes. That paper found that purely sequential difficulty stages cause forgetting, and that keeping easier tasks in the mix fixes it. Our analogue is to voxelize a small fraction of each batch at earlier warp strengths. This guards against losing far-range and large-object skills as the near-range grid sharpens. It is an ablation, not a core assumption.
Reproducible execution serves as the measurement instrument. The effects we chase may be about one mAP on nuScenes, where RSV's own gains span 0.7 to 4.3 points. Sparse-convolution kernels are sensitive to CUDA and driver versions, and DeepDIVA: A Highly-Functional Python Framework for Reproducible Experiments notes that deterministic flags neither remove all nondeterministic CUDA operations nor come free in speed. We therefore run every configuration inside one version-pinned container image, in the style of Bencher: Simple and Reproducible Benchmarking for Black-Box Optimization, which isolates benchmarks in pinned Docker/Singularity images. Every configuration gets at least five seeds, and seed-to-seed spread is reported as the noise floor any claimed gain must clear. Each run also writes an append-only log of configuration, schedule values and kernel versions, borrowed from the provenance trail in Audit-Grade Harness for Agent-driven Scientific Computation Workflows. The log means a reported gain can be traced to the exact environment that produced it.
Baselines
The naive reference is the unwarped detector: PointPillars, SECOND and Voxel-RCNN at their standard uniform voxel size. For PointPillars, the file's anchor figure is 86.10 AP on KITTI Moderate Car bird's-eye-view detection, from PointPillars: Fast Encoders for Object Detection from Point Clouds. Its purpose is to set the floor and confirm our reproduction matches published behavior.
The strongest competitor is fixed-warp RSV at its best setting. RSV reports that PointPillars gains 26.97 Pedestrian AP and 17.79 Cyclist AP on KITTI, and 0.7 to 4.3 mAP on nuScenes across baselines. It also uses about 397 MB of GPU memory, where a global 0.025 m grid needs about 982 MB. The file gives these gains only as deltas; no absolute RSV AP value was found on Lacuna, so we will reproduce absolute numbers in our container. To be fair, fixed RSV gets the same sweep over $s_{max}$ that we give the target $s^{*}_{max}$, the same epochs, data and augmentations, and the same number of tuning trials. We count as beating it only this: at matched GPU memory, the annealed schedule exceeds the best fixed $s_{max}$ on near-range (0–30 m) Pedestrian and Cyclist AP by more than twice the seed standard deviation, with no loss beyond noise at 50 m and farther.
The brute-force reference is a uniform fine grid at 0.05 m and 0.025 m. RSV reports that halving voxel size from 0.1 m to 0.05 m gains about 14 Pedestrian AP within 30 m and nothing beyond 50 m. This bounds how much near-range resolution alone is worth.
Three ablations isolate the key ingredient. The first is a two-stage schedule: train unwarped, then switch to full warp at the midpoint. The second is a curriculum over training samples by difficulty (easy, well-sampled objects first) with the warp held fixed, which tests whether any curriculum at all, rather than one over the representation, explains the gain. The third turns off the mixed-warp buffer.
Why not just…?
Why not just tune the fixed warp more carefully or train it longer? This is the most serious threat. RSV's large gains show fixed warps already train well at their chosen strength. Our $s_{max}$ sweep settles it directly. If fixed-warp accuracy rises monotonically, or plateaus without falling, across strengths we can afford in memory, there is no optimization ceiling and the proposal's main claim is false. We would report that negative result and not argue around it. We also give fixed RSV a longer training budget equal to the curriculum's longest run, so extra steps cannot explain a win.
Why not just pretrain on uniform voxels and fine-tune on the warped grid? That is a one-step curriculum, and it may be enough. The task-centric theory predicts it should lose to gradual annealing when the jump is large. Nothing in the file tests that in 3D perception, however. The two-stage ablation decides it. If two-stage matches continuous annealing within noise, the finding simplifies to "warm-start before warping." That is still useful, and cheaper, and we would recommend it.
Why not just let a feedback controller pick voxel size, as AVS-Net does? AVS-Net: Point sampling with adaptive voxel size for 3D scene understanding uses a proportional-integral controller to tune voxel size across downsampling layers toward a target sampling ratio. It reports 79.9 L1 and 71.7 L2 vehicle mAP on Waymo. Its controller aims at a point-count target, not at a training trajectory, and it needs per-dataset gains. It is also a different architecture, not a plug-in for grid detectors. We treat it as orthogonal rather than a head-to-head baseline. The one related probe we run is whether a fixed-strength controller is destabilized when our warp schedule shifts the point distribution. That is plausible but untested, and it is secondary.
Proposed Experiments
The primary study is on KITTI 3D detection with PointPillars, SECOND and Voxel-RCNN, with AP reported per class and per range bin (0–30, 30–50, 50+ m), plus GPU memory. We replicate on nuScenes, reporting mAP and the tracking-error metric AMOTP on which RSV reports a 24.6% change for CenterPoint-Pillar. The secondary study anneals NUC-Net's radial interval progression on SemanticKITTI segmentation, measured in mIoU. The file reports NUC-Net 5.8% mIoU above Cylinder3D and PolarNet at 54.3% mIoU.
The decisive checks, in order, are as follows. First, the fixed-warp sweep must show accuracy falling at high $s_{max}$; otherwise we stop. Second, annealing to a stronger target must beat the best fixed warp at equal memory by more than twice the seed spread, consistently across all three detectors. Third, gradual annealing must beat the two-stage switch, and there should be a flat region over the number of annealing stages, as moderate-separation theory predicts. Fourth, the gain must survive rerunning the whole grid on a second GPU generation using the same container image. If the ranking between methods flips across hardware, the effect is an artifact of kernel nondeterminism, not of the curriculum.
References
- Radial Scaling Voxelization for Accurate Small Object 3D Detection
- Curriculum DeepSDF
- A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
- FaceX-Zoo: A PyTorch Toolbox for Face Recognition
- NUC-Net: Non-uniform Cylindrical Partition Network for Efficient LiDAR Semantic Segmentation
- How does Multi-Task Training Affect Transformer In-Context Capabilities? Investigations with Function Classes
- DeepDIVA: A Highly-Functional Python Framework for Reproducible Experiments
- Bencher: Simple and Reproducible Benchmarking for Black-Box Optimization
- Audit-Grade Harness for Agent-driven Scientific Computation Workflows
- PointPillars: Fast Encoders for Object Detection From Point Clouds.
- AVS-Net: Point sampling with adaptive voxel size for 3D scene understanding