No Dense Tensors Needed: Fully Sparse Object Detection on Event-Camera Voxel Grids

The error forensics methodology is exemplary: the authors decompose all 9,673 failures across 119,459 test frames to show that 71% are localization near-misses (IoU∈(0,0.5)) rather than complete misses, and that recall reaches 95.4% at IoU≥0.40 versus 91.9% at IoU≥0.50. This demonstrates strong detection capability masked by box imprecision. The sparsity efficiency claims are well-quantified and practically significant: 858× GPU memory compression and 3,670× storage reduction versus dense 3D tensors, with active voxel counts increasing only 9% when resolution grows 2.25× (640² to 1280×720). The architecture is technically sound, adapting SEW-ResNet residuals and FCOS detection to the sparse domain via spconv.

The most significant issue is the counterintuitive finding that native 1280×720 resolution underperforms resized 640×640 by 2.0 mAP points (81.25% vs 83.22%), which the authors attribute to sparse kernel occupancy dropping from ~62% to ~30%. This suggests the sparse representation trades resolution for context density—a fundamental tension not fully resolved. The FPN's sparse transpose convolutions expand the active position set at each upsampling stage, partially eroding computational savings (acknowledged in Section 6.1 as 'FPN position expansion'). The model also shows early overfitting: validation loss minimum occurs at epoch 4 with subsequent divergence (Figure 5), raising questions about whether longer training with stronger regularization could close the gap. Finally, the 10-point mAP@50:95 deficit indicates severe regression precision issues at strict IoU thresholds that may limit applicability for tasks requiring tight bounding boxes.

The comparison to YOLOv11 on the FRED benchmark is fair and uses the same dataset splits. The evidence for the localization-precision hypothesis is strong: Table 3 shows mAP rises from 83.38% to 89.26% when relaxing IoU threshold from 0.50 to 0.40, and Figure 7b shows 43.1% of false negatives fall in IoU∈[0.4,0.5). The comparison to SAST and SFOD is accurate—the authors correctly note that prior sparse event detectors retain dense backbones or intermediate representations, while SparseVoxelDet maintains sparsity end-to-end. However, the paper lacks runtime/latency comparisons (acknowledged as confounded by implementation maturity), relying instead on theoretical position counts. The claim that sparse detectors scale with scene dynamics (O(N_events)) versus resolution (O(H×W)) is well-supported by Table 1 data showing only 9% voxel increase for 2.25× pixel growth.

Reproducibility is strong: all code, training configurations, evaluation scripts, and error forensics pipelines are publicly available on GitHub, and the FRED dataset is public. Training details are comprehensive: AdamW with lr=3×10⁻⁴, cosine schedule with 5k-step warmup, batch size 2, mixed-precision FP16, EMA decay 0.9997, focal loss (α=0.25,γ=2.0), and specific augmentation (horizontal flip, polarity inversion, event dropout, random scaling). The authors report multi-seed results (seeds 42 and 123) with low variance (±0.16 pp at 640², ±0.14 pp at native), confirming training stability. Single-GPU training on RTX 3090 makes the work accessible without large compute infrastructure. The error forensics methodology (Section 4.6) is described in sufficient detail to be reusable for other single-class detectors.