吞下17亿图片，Meta最强巨兽DINOv3开源，重新定义CV天花板

【导读】无需人工标注，吞下17亿张图片，Meta用自监督学习炼出「视觉全能王」！NASA已将它送上火星，医疗、卫星、自动驾驶领域集体沸腾。 | Task | DINO | DINOv2 | DINOv3 | SigLIP 2 | PE | | --- | --- | --- | --- | --- | --- | | Benchmark | ViT-B/8 | ViT-g/14 | ViT-7B/16 | VIT-g-OPT/16 | ViT-G/14 | | | 0.09B | 1.1B | 7B | 1.8B | 1.98 | | Segmentation ADE-20k | 31.8 | 49.5 | 55.9 | 42.7 | 38.9 | | Depth estimation NYU I | 0.537 | 0.372 | 0.309 | 0.494 | 0.436 | | Video tracking DAVIS | 68.7 | 76.6 | 83.3 | 62.9 | 49.8 | | Instance retrieval Met | 17.1 | 44.6 | 55.4 | 13.9 | 1 ...