ZoeDepth NYU KITTI

ZoeDepth (fine-tuned on NYU and KITTI) is a monocular depth estimation model that outputs metric (absolute) depth values. It extends the DPT architecture with a novel metric bins module, enabling zero-shot transfer to unseen domains while preserving scale accuracy.

Architecture and Training

The model variant ZoeD-M12-NK is pre-trained on 12 datasets for relative depth and fine-tuned jointly on NYU Depth v2 and KITTI for metric depth. During inference, an internal latent classifier routes each input to the appropriate domain-specific head, allowing it to handle both indoor and outdoor scenes without retraining.

Benchmark Performance

On the NYU Depth v2 indoor benchmark, ZoeDepth achieves a 21% improvement in relative absolute error (REL) over the prior state of the art. It is the first model to jointly train on multiple metric depth datasets (NYU and KITTI) without significant performance drop, and demonstrates zero-shot generalization to eight unseen datasets spanning indoor and outdoor environments.

Maintenance Status

The original implementation repository is no longer actively maintained by Intel. As a hosted API on gigarouter, the model is served without requiring local installation or maintenance.