Predictive Self-Conditioning for Accurate and Calibrated LiDAR 3D Object Detection

Two-stage LiDAR detectors commit their bird’s-eye-view (BEV) feature map before the detection head runs, so the evidence the first detection pass produces — where objects were found, and how reliable each detection is — can never improve the representation it was computed from. We close this gap with Predictive Self-Conditioning (PSC): a within-frame, feature-level feedback mechanism that converts Pass-1 proposals into per-class, confidence-weighted Gaussian heatmaps and injects them as a BEV feature residual before Pass-2, letting a detector refine its own representation from its own first-pass evidence, with no teacher and no temporal memory.

PSC is the core of our detector PSC-PV-RCNN, flanked by two optional modules: Density-Gated BEV Fusion (DGBF), enriching the features PSC conditions on, and a Region-Aware conditional variational autoencoder (RA-CVAE), a per-proposal uncertainty gate targeting calibration. PSC attaches to any detector that materialises a dense BEV map, and transfers across multiple backbones and head families: against matched baselines, it achieves +2.48 Moderate mAP on KITTI (+3.95 pedestrian, +3.44 cyclist) and +1.15 mean LEVEL 2 heading-weighted AP (APH) on the Waymo Open Dataset. In addition, the LiDAR-only Region-Aware CVAE lowers Expected Calibration Error (ECE) by 12.7% without requiring camera inputs.

Recommended citation: Sami Sadat, Md. Irtiza Hossain, Shreya Adrita Banik, Md Zia Uddin, Md. Golam Rabiul Alam. (2027). "Predictive Self-Conditioning for Accurate and Calibrated LiDAR 3D Object Detection." IEEE/CVF Winter Conference on Applications of Computer Vision (WACV, CORE A). Accepted.