LingBot-Vision
A boundary-centric self-supervised vision foundation model, built for dense spatial perception
Overview
Dense spatial perception is essential for physical intelligence, yet modern visual foundation models tend to prioritize semantic invariance at the expense of detailed spatial understanding. LingBot‑Vision studies vision pretraining through a boundary‑centric lens: boundaries and shape discontinuities offer essential cues for perceiving geometric properties.
At its core is masked boundary modeling: without pretrained backbones, human annotations, or external edge detectors, the model learns sub‑pixel boundary representations online and uses the discovered boundary‑bearing tokens as masked targets for dense visual token learning. Scaled to a 1B‑parameter ViT‑g/16, LingBot‑Vision delivers leading dense perception against strong baselines such as DINOv3, and powers LingBot‑Depth 2.0.
Boundary‑Faithful Dense Features
We map the top three PCA components of the frozen patch features to RGB, with no fine‑tuning. LingBot‑Vision resolves objects as coherent, sharply bounded regions: interiors stay smooth while feature transitions land exactly on object contours, with none of the per‑token speckle or blocky noise of prior backbones. Select a scene to compare the input video with its dense feature map.
High‑Resolution Feature Inference
The same frozen features scale directly to high‑resolution inference: on a 4096-token grid, the PCA maps preserve whiskers, cables, floor‑tile patterns and other thin structures. Each pair shows the input photo (left) and its dense feature map (right).
Boundary Tokens
During pretraining, LingBot‑Vision discovers boundary‑bearing tokens online, with no annotations and no external edge detectors. The boundary tokens tightly trace object silhouettes, whiskers, cables and screen edges, and each token carries a sub‑pixel boundary field. Querying any boundary token by cosine similarity of the frozen features lights up the corresponding structure, and nothing else.
Another example: input / boundary tokens / feature PCA
Tracking Boundary Tokens Through Time
Boundary tokens behave as stable, trackable entities. Queries selected in the first frame are tracked frame by frame using only cosine similarity of the frozen patch features, with no fine‑tuning and no temporal supervision: queries on static structures stay put, while object queries follow large pose changes, and the similarity responses remain selective over the full clip.
Powering LingBot‑Depth 2.0
LingBot‑Depth formulates depth completion as masked depth modeling (MDM), and the single external dependency of the recipe is the encoder initialization. LingBot‑Depth 2.0 swaps the DINOv2 encoder for LingBot‑Vision, at both ViT‑L/16 and ViT‑g/16, and scales the curated RGB‑D training corpus from 3M to 150M samples.
The two tweaks compound: under the identical pipeline, the LingBot‑Vision initialization is the strongest starting point at both scales (e.g. 0.094 vs 0.152 RMSE against DINOv2 on block‑masked DIODE‑Indoor at ViT‑L), and its lead widens rather than washes out as data grows. LingBot‑Depth 2.0 sets the leading performance on 14 benchmarks spanning depth patterns and camera types, halving the block‑masked DIODE‑Indoor RMSE of version 1.0 (0.132 to 0.062) and excelling on transparent‑object captures, the classic failure case of active depth sensing. Depth completion concentrates its uncertainty around object boundaries and material transitions, exactly the structure that the boundary‑anchored features of LingBot‑Vision encode.
Encoder initialization: block‑masked DIODE‑Indoor RMSE (lower is better)
Depth completion, method by method
Raw sensor depth is missing exactly on the hardest surfaces: transparent and reflective objects return no measurements at all. Pick a scene and switch between methods to compare depth completion systems on real camera captures.
Drag to rotate, scroll to zoom. The viewpoint persists when switching methods, so the point‑cloud geometry of each method can be compared directly.
Benchmark Results
All results use frozen features: a single linear layer for dense probing and training‑free label propagation for video, so performance reflects the representation rather than the readout.
Linear probing: depth estimation (RMSE, lower is better)
Linear probing: semantic segmentation (mIoU, higher is better)
Training‑free video object segmentation (J&F, higher is better)
Distilled Model Family
The giant teacher is distilled into ViT‑L, ViT‑B and ViT‑S students, and the dense advantage transfers essentially intact: at every deployment budget, the LingBot‑Vision student offers the best depth accuracy in its class, with linear‑probe classification within a point of the best competitor. The 0.3B ViT‑L student matches the 7B DINOv3 on NYUv2 (0.310 vs 0.309) with roughly 23x fewer parameters.
| Size | Model | IN1k Linear | IN1k KNN | NYUv2 RMSE ↓ | KITTI RMSE ↓ | ADE20K | VOC12 | Cityscapes |
|---|---|---|---|---|---|---|---|---|
| ViT-L | DINOv2 | 86.43 | 83.82 | 0.411 | 3.243 | 48.63 | 83.54 | 74.51 |
| DINOv3 | 87.31 | 85.27 | 0.351 | 2.643 | 55.00 | 88.11 | 79.28 | |
| SigLIP2 | 85.67 | 82.91 | 0.516 | 3.641 | 42.99 | 75.55 | 62.51 | |
| Franca | 82.50 | 71.70 | 0.425 | 3.613 | 45.93 | 84.42 | 72.42 | |
| V-JEPA2 | -- | -- | 0.346 | 2.617 | 46.29 | 84.81 | 71.58 | |
| LingBot-Vision | 86.38 | 83.62 | 0.310 | 2.574 | 52.75 | 87.40 | 78.75 | |
| ViT-B | DINOv2 | 84.27 | 82.17 | 0.429 | 3.416 | 48.25 | 83.81 | 74.19 |
| DINOv3 | 84.79 | 83.21 | 0.371 | 2.826 | 51.74 | 88.20 | 77.30 | |
| SigLIP2 | 82.47 | 79.47 | 0.539 | 3.847 | 40.49 | 73.50 | 60.12 | |
| Franca | 78.45 | 63.45 | 0.462 | 3.790 | 42.69 | 82.79 | 71.65 | |
| V-JEPA2 | -- | -- | 0.410 | 2.883 | 41.38 | 78.00 | 64.99 | |
| LingBot-Vision | 85.05 | 82.32 | 0.339 | 2.793 | 51.44 | 87.10 | 77.17 | |
| ViT-S | DINOv2 | 80.76 | 79.11 | 0.447 | 3.568 | 44.86 | 82.34 | 71.44 |
| DINOv3 | 80.37 | 79.33 | 0.405 | 2.851 | 46.51 | 84.30 | 73.38 | |
| LingBot-Vision | 82.22 | 78.98 | 0.383 | 3.784 | 47.01 | 84.17 | 64.34 |