LingBot-Vision Surpasses DINOv3 Performance with Boundary Modeling

Instead of random patch masking, LingBot-Vision uses a Masked Boundary Modeling approach that forces the Student model to reconstruct the Boundary field predicted by the Teacher model.

The key features and performance are as follows:

• Efficient training: With only 161 million images, less than a third of DINOv3's data, it recorded 0.296 RMSE on the NYUv2 benchmark, surpassing the 7B-scale DINOv3 (0.309).
• Core design: The boundary field is reconstructed as per-pixel categorical distributions to prevent model collapse during training, and a-contrario validation is performed on decoded segments to manage training quality.
• Scalability: It shows high parameter efficiency, with the ViT-L (0.3B) model achieving performance similar to the 7B model, and also demonstrated superior initialization performance compared to DINOv2 in data scaling curves.

However, it showed somewhat lower performance compared to the DINOv3 family on the ImageNet classification, ADE20K, and KITTI benchmarks.