Instance-Level Video Depth in Groups Beyond Occlusions

a) Objects in a Group (b) Depth Anything V2 [53] (c) Zoedepth [4] (d) Instance Mask (e) Ours Figure 1. A typical scenario involving a complex dynamic group that cannot be tackled by existing monocular depth estimation methods.

We contribute the first Group Instance Depth Dataset and a novel instance-aware depth estimation framework. Our method effectively captures instance-wise depth relationships, mitigating occlusion ambiguities and improving depth accuracy in dynamic multi-object scenes.