Question about the impressive transparent glass depth results

[Chen94yue]

Hi Lingbot-Depth team,

First of all, thank you for releasing this great work. The results of Lingbot-Depth are very impressive, especially in challenging indoor scenes with transparent glass, glass doors, or reflective surfaces.

I noticed that the predicted depth around transparent glass regions looks surprisingly robust. This is a very difficult hard case for RGB-D depth completion, since the raw depth from sensors can be missing, noisy, or even penetrate through the glass and hit the background instead.

I would like to better understand what mainly contributes to this strong performance on transparent glass scenes:

1. Is the improvement mainly due to the Lingbot-Vision backbone and its stronger visual/structural representation?
2. Or did the training data include specific transparent-glass / glass-door / reflective-surface scenes?
3. If special data was used, was it collected with dense ground-truth depth for glass surfaces, or was it handled through pseudo labels, synthetic data, or other supervision strategies?
4. Are there any recommended practices for adapting Lingbot-Depth to transparent glass hard cases in custom indoor robot scenarios?

I am particularly interested in whether the model learns this capability mostly from the general-purpose vision backbone, or whether transparent/glass-specific data is still necessary for robust performance.

Thanks again for the excellent work!

[cherubicXN]

Hi, thank you for your kind words and thoughtful questions. Let me address them one by one:
The improvement mainly comes from our new pretrained backbone, together with mid-training on more than 140 million diverse RGB-D samples.

Yes, our RGB-D corpus is highly diverse and includes many challenging cases, including transparent glass, glass doors, and reflective surfaces.

Obtaining perfect depth ground truth at this scale is not feasible. We therefore use pseudo-labels generated with a strong depth-prior model—LingBot-Depth itself—as part of a robust data pipeline designed to support various RGB-D cameras. We also incorporate synthetic data to help resolve ambiguities in challenging regions.

For adapting the model to custom indoor robotic scenarios, a cost-effective approach is to post-train it using high-quality synthetic data that closely matches the target environment and sensor characteristics.

Overall, our depth model benefits from the best of both worlds: strong, general-purpose visual representations learned during pretraining and high-quality, data-driven mid-training tailored to depth perception. Robust performance on transparent and reflective surfaces comes from the combination of these factors rather than from either one alone.