Stereo Depth Lab
An open-source prototype project that converts images and short videos into stereo pairs. It uses a relative depth model and a Gaussian rendering pipeline that carries points across frames through optical flow, confidence, decay and scene-change resets.
Timeline (last 90 days)
Perfect depth in a single frame does not guarantee that the geometry of depth-to-stereo video remains temporally stable; frame-by-frame comparison misses contour jitter, the defect that dominates the viewing experience.
Stereo Depth Lab is a local research prototype that converts images and short video clips into stereo pairs, using Depth Anything V2 Small by default as the relative depth estimation model and a lightweight Gaussian rendering path with optical flow, confidence, decay, and scene-change reset.
Depth Anything V2 is a relative depth model, so when Stereo Depth Lab uses it to generate stereo pairs, the interocular distance and convergence parameters are not derived directly from depth values, but are set by the user and combined with scaling.
Public cases document the confidence, decay, and scene-change reset mechanisms of the Stereo Depth Lab prototype, but do not provide benchmark evidence proving they can solve all situations; the occlusion ghosting and other issues described in the text are illustrative failure cases.
All Facts (4)
Perfect depth in a single frame does not guarantee that the geometry of depth-to-stereo video remains temporally stable; frame-by-frame comparison misses contour jitter, the defect that dominates the viewing experience.
50%UnverifiedStereo Depth Lab is a local research prototype that converts images and short video clips into stereo pairs, using Depth Anything V2 Small by default as the relative depth estimation model and a lightweight Gaussian rendering path with optical flow, confidence, decay, and scene-change reset.
50%UnverifiedDepth Anything V2 is a relative depth model, so when Stereo Depth Lab uses it to generate stereo pairs, the interocular distance and convergence parameters are not derived directly from depth values, but are set by the user and combined with scaling.
50%UnverifiedPublic cases document the confidence, decay, and scene-change reset mechanisms of the Stereo Depth Lab prototype, but do not provide benchmark evidence proving they can solve all situations; the occlusion ghosting and other issues described in the text are illustrative failure cases.
50%