Modern depth predictors adopt physically unsupported edge cues as real geometry, causing global structural hallucinations that local repair cannot fix.
hub Canonical reference
ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth
Canonical reference. 70% of citing Pith papers cite this work as background.
74
Pith papers citing it
173
external citations · Pith
Background
70% of classified citations
abstract
This paper tackles the problem of depth estimation from a single image. Existing work either focuses on generalization performance disregarding metric scale, i.e. relative depth estimation, or state-of-the-art results on specific datasets, i.e. metric depth estimation. We propose the first approach that combines both worlds, leading to a model with excellent generalization performance while maintaining metric scale. Our flagship model, ZoeD-M12-NK, is pre-trained on 12 datasets using relative depth and fine-tuned on two datasets using metric depth. We use a lightweight head with a novel bin adjustment design called metric bins module for each domain. During inference, each input image is automatically routed to the appropriate head using a latent classifier. Our framework admits multiple configurations depending on the datasets used for relative depth pre-training and metric fine-tuning. Without pre-training, we can already significantly improve the state of the art (SOTA) on the NYU Depth v2 indoor dataset. Pre-training on twelve datasets and fine-tuning on the NYU Depth v2 indoor dataset, we can further improve SOTA for a total of 21% in terms of relative absolute error (REL). Finally, ZoeD-M12-NK is the first model that can jointly train on multiple datasets (NYU Depth v2 and KITTI) without a significant drop in performance and achieve unprecedented zero-shot generalization performance to eight unseen datasets from both indoor and outdoor domains. The code and pre-trained models are publicly available at https://github.com/isl-org/ZoeDepth .
hub tools
citation-role summary
background 7
method 2
baseline 1
citation-polarity summary
representative citing papers
Introduces MultiDepth-3k benchmark revealing diverse layer preferences across depth models on ambiguous scenes, with Laplacian Visual Prompting altering outputs for some frozen models and best pair reaching 75.5% ML-SRA.
DepthMaster unifies metric monocular depth estimation for perspective and panoramic images by patching panoramas into perspective views, adding a consistency loss and virtual cameras, and training mostly on perspective data to reach SOTA zero-shot results on 13 datasets.
MetricScenes dataset from web photos and stereo imagery, plus a two-stage Poisson depth completion method, allows fine-tuning MoGe-2 to mitigate scale-collapse in metric monocular geometry while preserving benchmark performance.
TROPHIES introduces a unified framework for human-scene-camera reconstruction from multi-view videos, achieving globally aligned and physically plausible 4D outputs on EgoHuman and EgoExo4D.
SeeGroup formulates per-pixel multi-layer depth as a point process with permutation-invariant likelihood to support arbitrary groupings, raising quadruplet relative depth accuracy from 61.34% to 70.09% on the LayeredDepth benchmark.
GAMSI is a dual-pathway Geometry-Aware MLLM using Metric-Structure Decoupled Queries and Expert-Guided Visual Grounding on RGB inputs alone, trained on a new 152k-sample MTS dataset to reach SOTA on seven spatial benchmarks.
WideDepth supplies the first millimeter-accurate indoor fisheye depth benchmark together with a stereo generation pipeline and model adaptation technique.
H-Flow learns dense human scene flow from monocular video via joint pose and depth prediction in a multi-head transformer, using physics-inspired geometric and biomechanical priors for self-supervision, and introduces the DynAct4D synthetic benchmark.
Depth2Pose is a new evaluation framework for monocular depth estimators that uses relative camera pose accuracy as a task-driven proxy and introduces the D2P dataset of challenging out-of-distribution scenes.
LAMP tracks 3D human motion from moving multi-camera headsets by converting 2D detections to a unified metric 3D world frame via device localization and fitting with an end-to-end spatio-temporal transformer.
Dual-pixel defocus blur enables absolute scale estimation in SfM without reference objects or calibration.
A video generation approach conditions a base model with multi-scale 3D latent features and a cross-attention adapter to produce geometrically realistic and consistent orbital videos from one image.
LiftFormer transforms monocular depth prediction into depth-oriented geometric and edge-aware subspace representations via lifting and frame theory, achieving state-of-the-art results on standard datasets.
EndoVGGT uses a dynamic DeGAT graph attention module to improve depth estimation and non-rigid 3D reconstruction in surgery, reporting 24.6% PSNR and 9.1% SSIM gains on SCARED with zero-shot generalization to new domains.
VGGT-360 delivers geometry-consistent zero-shot panoramic depth by converting panoramas into multi-view 3D reconstructions via VGGT models and three plug-and-play correction modules, then reprojecting the result.
RAD retrieves semantically similar RGB-D context samples for low-confidence regions and fuses them via matched cross-attention to cut relative absolute depth error by 29.2% on NYU Depth v2 underrepresented classes while staying competitive on standard benchmarks.
URF-GS creates a single radiation field from visual and wireless observations via 3D Gaussian splatting to predict radio signals at any location and configuration with higher accuracy and fewer samples than prior NeRF approaches.
Proposes the first light field-LiDAR semantic segmentation dataset and the Mlpfseg network, which improves mIoU by 1.71 over image-only and 2.38 over point-cloud-only baselines via feature completion and depth perception modules.
Materialist performs single-image inverse rendering via neural-initialized progressive differentiable rendering to enable physically consistent material editing, object insertion, relighting, and transparency edits without full scene geometry.
3D-VLA is a new embodied foundation model that uses a 3D LLM plus aligned diffusion models to generate future images and point clouds for improved reasoning and action planning in 3D environments.
SynCity 3000 generates large, coherent 3D scenes from text by fine-tuning an image-to-3D diffusion model to operate convolutionally on overlapping windows, trained on procedurally generated synthetic scene data.
A 68K-pair aerial UAV depth benchmark shows standard monocular metric-depth models collapse in the air, and LoRA fine-tuning on it (MoGe2-Aerial) restores accurate metric depth on both aerial and ground scenes.
Error-Conditioned Neural Solvers improve PDE prediction accuracy by using the residual field as network input for learned corrections, outperforming residual-minimization methods by up to 10x on turbulent flows and generalizing better under distribution shifts.
citing papers explorer
-
SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model
SpatialVLA adds 3D-aware position encoding and adaptive discretized action grids to visual-language-action models, enabling strong zero-shot performance and fine-tuning on new robot setups after pre-training on 1.1 million real-world episodes.
-
Step1X-Edit: A Practical Framework for General Image Editing
Step1X-Edit integrates a multimodal LLM with a diffusion decoder, trained on a custom high-quality dataset, to deliver image editing performance that surpasses open-source baselines and approaches proprietary models on the new GEdit-Bench.