ar-ray-code/lingbot-depth-trt

lingbot-depth-trt is a set of tools that help run an existing AI model, called LingBot-Depth, faster on NVIDIA hardware, along with a live demo that shows the model working with a real depth camera. LingBot-Depth itself, a model for estimating depth, meaning how far away things are, from camera images, is a separate project this repository depends on rather than including directly, anyone using this tool is asked to also credit and refer to that original project. The main job of this repository is converting the LingBot-Depth model into a format called TensorRT, which is NVIDIA's own format for running AI models more efficiently on their graphics cards. To do this conversion, a person first downloads a pretrained version of the model from Hugging Face, a site that hosts AI models, and then runs a conversion script that produces a TensorRT engine file sized for a fixed image resolution, 640 by 480 pixels by default, though other sizes can be requested. Once converted, the project offers a live demo that connects to a depth camera, either an Intel RealSense D405 or an Orbbec Gemini 335, and shows the depth output either in a regular window or in a 3D point cloud viewer that runs in a web browser. In the browser viewer, raw depth data and depth data cleaned up by the model are shown side by side, with synchronized camera controls, so a person can click a point in the scene and see its exact 3D coordinates and distance. There is also a separate demo focused on tracking the camera's own movement over time, using both the depth data and the camera's built-in motion sensor, again viewable in the browser. The setup instructions are specific and demanding: the project was built and tested on Ubuntu 26.04 with a particular NVIDIA driver version and an RTX 3060 graphics card, using a package manager called uv rather than the more common pip. Support for each supported camera brand is installed separately as an optional add-on. The README does not state a license for this code, though it does provide citations for the two research papers the underlying model is based on.