What is mda?

MDA is the official code release for a research paper about estimating depth from images and video, meaning figuring out how far away each part of a scene is from the camera. The paper is called Modeling Depth Ambiguity, and it comes from researchers named Siyuan Bian, Congrong Xu, and Jun Gao. Most depth estimation models predict a single depth value for every pixel in an image. The problem the authors identify is that at the edges of objects, a pixel often covers both the foreground object and the background behind it, so a single depth number cannot describe it well. Averaging the two distances creates what the README calls a flying point, a false surface that appears to float between the real foreground and background and corrupts the resulting shape. Instead of one value per pixel, MDA predicts a small set of possible depth values along with probabilities for each, then chooses the most likely one rather than blending them together. According to the README this largely removes flying points, holds up better when the input image is blurry, adds very little extra computing cost, and can be attached to two existing depth backbones called DA3 and VGGT. The repository includes code to run the trained models on your own images, videos, or folders of photos, with an interactive point cloud viewer that opens automatically after each run. Two pretrained checkpoints are available for download from Hugging Face, one built on the DA3 backbone and one on VGGT. It also includes the training code, built on the Hydra and Lightning frameworks, along with evaluation scripts that reproduce the benchmark results reported in the paper across several depth and video datasets. This is a research codebase aimed at people working on computer vision and 3D reconstruction, not a general consumer tool. Getting it running requires a CUDA capable GPU, a Python environment set up with PyTorch, and for the video demo, ffmpeg installed on the system. The README documents installation, checkpoint downloads, and training and evaluation commands in detail.