LingBot-Vision: masked boundary modeling for self-supervised pretraining (0.296 NYUv2 linear-probe RMSE at 1.1B vs 0.309 for DINOv3-7B, trails on ImageNet); weights in 4 sizes[R]

LingBot-Vision introduces a new method for self-supervised pretraining, focusing on masked boundary modeling. This approach involves a teacher predicting a dense boundary field, which is then used to force boundary-bearing tokens into the student's mask. The student must reconstruct the regions that cannot be inferred by copying context.

The boundary targets in LingBot-Vision come from the teacher itself, rather than labels or an external edge detector. This allows for a more nuanced understanding of boundary structures.

The results of LingBot-Vision are promising, with a 0.296 NYUv2 linear-probe RMSE, surpassing the 0.309 achieved by DINOv3-7B. However, it trails on ImageNet.

The LingBot-Vision model is available in four sizes, providing flexibility for various applications.

This development has significant implications for the field of computer vision, as it demonstrates the potential of masked boundary modeling for self-supervised pretraining.

Source: LingBot-Vision: masked boundary modeling for self-supervised pretraining (0.296 NYUv2 linear-probe RMSE at 1.1B vs 0.309 for DINOv3-7B, trails on ImageNet); weights in 4 sizes[R]. Read the full piece at the source.