Vision Pretraining for Dense Spatial Perception

Vision Pretraining for Dense Spatial Perception proposes a boundary-centric self-supervised objective that treats boundaries as native learning signals rather than downstream outputs. Instead of random masking, the method uses masked boundary modeling: a teacher Vision Transformer discovers boundary-bearing tokens online, forces them into the student’s masked set, and supervises them with a categorical boundary-field loss over discretized distance and orientation bins. This reparameterization stabilizes dense self-distillation and enables parameter-free a-contrario validation of candidate line segments, so unsupported structure never becomes a training target. Scaled to LingBot-Vision, a 1B Vision Transformer trained on 160.75M curated images, the approach matches or surpasses foundation models up to 7B parameters on dense spatial perception. On NYU-Depth v2 linear probing, LingBot-Vision reaches 0.296 RMSE, ahead of DINOv3’s 0.309, while on DAVIS-2017 and YouTube-VOS it posts 70.0 and 73.5 J&F-Mean, respectively, near the strongest distilled baselines. The paper also shows that a 0.3B student matches the 7B DINOv3 on NYU-Depth v2 with roughly 23× fewer parameters. These pretrained encoders transfer directly to LingBot-Depth 2.0, where replacing the initialization and scaling RGB-D training data from 3M to 150M samples yields leading depth completion on 14 benchmarks, especially on hard transparent and reflective scenes.