DRA-Pillars: a controlled study of density- and range-motivated design choices for pillar-based LiDAR 3D detection

LiDAR point sparsity and range-dependent sampling motivate density-aware and range-motivated design choices in pillar-based 3D detection. Such choices are often assessed through isolated or single-run ablations, making it difficult to determine whether apparent improvements are stable across random seeds and whether component additions yield consistent improvements under a common evaluation protocol. We study these questions through DRA-Pillars, a pillar-based framework that operationalizes density-aware coordinate attention (DACA), an adaptive gated feature pyramid (AGFF), range-adaptive (RA) supervision, and a decoupled prediction head. The controlled evaluation uses the KITTI Car class with a custom 80/20 split comprising 5984 training and 1496 validation frames, three paired random seeds, a common evaluation protocol, and APR40 for Car 3D detection. Five primary configurations (E00 to E04) characterize cumulative component effects, and the additional E05 configuration supports the RA/head design analysis. Under this protocol, DACA does not produce a consistent positive effect. AGFF produces a modest positive mean increment, but its effect remains heterogeneous across seeds. RA-related supervision produces a negative mean effect under the tested standard-head configuration and is associated with increased variability. The decoupled-head package raises the three-seed mean relative to E03, although the increase is not reproduced by every seed. For moderate APR40, the means are 84.582 for E00, 82.808 for E03, 83.938 for E04, and 85.689 for E05. Among the six evaluated configurations, E05 yielded the highest three-seed mean APR40 across easy, moderate, and hard. The four-cell E02/E03/E04/E05 RA-by-decoupled-head interaction remains inconclusive, and the E04 three-seed mean does not exceed the E00 mean under the common evaluation protocol. These results indicate that density- and range-motivated design choices can have heterogeneous and nonmonotone effects, so single-run or mean-only component conclusions can obscure seed-level instability. The claims are limited to the tested KITTI Car protocol.