LingBot-Vision系列模型发布

LingBot-Vision is now on ModelScope: a boundary-first vision model series for cleaner depth, segmentation, and video tracking. 🚀 Four checkpoints: LingBot-Vision-Giant/Large/Base/Small 🤖 https://t.co/GXPI6FfWBX 📄 https://t.co/5wlllraB6x 🏆 Dense perception: the 1B ViT-g model reaches 0.296 NYUv2 depth RMSE, ahead of 7B DINOv3 at 0.309; it also matches Cityscapes 79.6 and leads VOC12 at 87.5. ⚡ Four sizes: ViT-g teacher plus ViT-L, ViT-B, and ViT-S students. The 0.3B ViT-L matches 7B DINOv3 on NYUv2, 0.310 vs 0.309, with roughly 23x fewer parameters. 🔬 Tech frame: masked boundary modeling learns boundary-aware visual features from scratch, without human annotations or external edge detectors. 🧠 Downstream backbone: powers LingBot-Depth 2.0, helping scale RGB-D training from 3M to 150M samples and improve depth completion across 14 benchmarks. License: Apache 2.0.