This work presents Depth Anything, a highly practical solution for robust monocular depth estimation. Without pursuing novel technical modules, we aim to build a simple yet powerful foundation model dealing with any images under any circumstances. To this end, we scale up the dataset by designing a data engine to collect and automatically annotate large-scale unlabeled images (approximately 62M). Our data scaling-up strategy is based on two core designs. First, we create a more challenging optimization target by leveraging data augmentation tools to compel the model to actively seek extra visual knowledge from the unlabeled data. Second, we develop an auxiliary supervision from pre-trained encoders to enforce the inheritance of rich semantic priors. Combined with 1.5M labeled images and 62M unlabeled images, our model shows impressive zero-shot capability on six public datasets and randomly captured photos. When fine-tuned with metric depth information from NYUv2 and KITTI, it sets new state-of-the-art results. Our approach also yields better depth-conditioned ControlNet.

Monocular Depth Estimation (MDE) is a fundamental problem with broad applications in robotics, autonomous driving, virtual reality, and many more fields. Similar to what has happened in NLP and other computer vision tasks, MDE also requires a foundation model to estimate depth information from a single image. However, building such a foundation model for MDE has remained challenging because obtaining tens of millions of precise depth labels is prohibitively expensive. While large-scale labeled datasets have driven progress in image classification and object detection, depth annotation demands specialized sensors (e.g., LiDAR, structured light) and careful calibration, severely limiting data scale.

MiDaS represented pioneering work in multi-dataset joint training for zero-shot relative depth estimation, but it suffered from limited data coverage and could not scale beyond the available labeled datasets. This work proposes to leverage large-scale monocular unlabeled images, which offer three key advantages: (1) simplicity and low cost of acquisition, (2) diversity covering broader scene ranges, and (3) ease of annotation using pre-trained models. We design a data engine to automatically generate depth annotations for unlabeled images, enabling data scaling-up to arbitrary scale. We collect 62M diverse images from eight large-scale public datasets including SA-1B, Open Images, and BDD100K, while using 1.5M labeled images from six public datasets to train an initial teacher model as the annotation tool.

Our initial attempt of directly combining labeled and pseudo-labeled images for training failed to bring noticeable improvements, since the labeled images were already sufficient and the student model could not learn extra knowledge from pseudo labels alone. To address this, we introduce two key strategies. First, we challenge the student model with stronger perturbations — including color distortions (color jittering, Gaussian blurring) and spatial distortion via CutMix — when learning from unlabeled images, forcing the model to acquire more robust representations under various distortions. Second, rather than assigning semantic segmentation labels using models like RAM, GroundingDINO, and HQ-SAM (which failed due to information loss in discrete class space), we preserve rich semantic information through a feature alignment loss from a frozen DINOv2 encoder, maintaining semantic priors while enhancing depth estimation performance.

Monocular Depth Estimation has evolved significantly from early approaches relying on handcrafted features and traditional computer vision methods. Deep learning methods revolutionized the field through carefully annotated datasets, with successive improvements coming through classification-based regression, additional geometric priors, and better objective functions. However, generalization to unseen domains remained a persistent challenge, as models trained on specific datasets often failed when applied to images from different environments, lighting conditions, or camera intrinsics.

Zero-shot depth estimation aims to train MDE models on diverse datasets enabling prediction on any image. MiDaS pioneered multi-dataset joint training using affine-invariant loss to ignore depth scale and shift variations across different datasets, providing relative depth information. Recent work explored metric depth estimation with zero-shot transfer, though with observed generalization trade-offs. On the side of leveraging unlabeled data, semi-supervised learning typically assumes limited labeled images. This work addresses a more challenging scenario: sufficient labeled images already exist, but even larger-scale unlabeled data can further push the performance boundary.

The approach uses a labeled set D^l = {(x_i, d_i)}_i=1^M and an unlabeled set D^u = {u_i}_i=1^N. A teacher model T is first learned from D^l, then assigns pseudo labels to D^u, and finally a student model S trains on the combined labeled and pseudo-labeled sets. The depth value is first transformed into disparity space by d = 1/t and then normalized to 0~1 on each depth map. The method adopts affine-invariant loss enabling multi-dataset joint training: L_l = (1/HW) Σ ρ(d̂_i, d_i), where ρ represents the affine-invariant mean absolute error with scale-and-shift alignment: d̂_i = (d_i − t(d)) / s(d), with t(d) = median(d) and s(d) = mean absolute deviation. Training employed 1.5M labeled images from six datasets (BlendedMVS, DIML, HRWSI, IRS, MegaDepth, TartanAir) while excluding NYUv2 and KITTI for zero-shot evaluation integrity. DINOv2 pre-trained weights initialized the encoders.

The pseudo-labeled set is created as D̂^u = {(u_i, T(u_i)) | u_i ∈ D^u}. Initial naive self-training that simply combined labeled and pseudo-labeled images failed to bring improvements, hypothesized to stem from the limited additional knowledge when teacher and student share the same architecture and pre-training. The solution introduces strong perturbations: color distortions (color jittering, Gaussian blurring) and spatial distortion via CutMix. CutMix interpolates random image pairs: u_ab = u_a ⊙ M + u_b ⊙ (1 − M), where M is a binary mask with a rectangle region set to 1. The unlabeled loss combines masked and unmasked regions: L_u = (ΣM / HW) L_u^M + (Σ(1−M) / HW) L_u^1−M. CutMix is applied with 50% probability. Crucially, unlabeled images fed to the teacher T for pseudo-labeling remain clean without any distortions.

Initial attempts to assign semantic segmentation labels to unlabeled images using RAM, GroundingDINO, and HQ-SAM models producing a 4K-class space failed to boost performance, attributed to information loss in the discrete class space. Instead, the method leverages DINOv2's strong semantic capabilities via a feature alignment loss: L_feat = 1 − (1/HW) Σ cos(f_i, f'_i), where cos(·, ·) measures cosine similarity between the depth model features f and the frozen DINOv2 encoder features f'. A critical observation is that semantic encoders like DINOv2 tend to produce similar features for different parts of an object (e.g., car front and rear), which conflicts with depth variation within the same object. To address this, a tolerance margin α is implemented: when cosine similarity already exceeds α, those pixels are excluded from L_feat, allowing both semantic-aware and part-level discriminative representations. The final loss combines three components: L_l, L_u, and L_feat via simple averaging.

Comprehensive validation is performed across six unseen datasets: KITTI, NYUv2, Sintel, DDAD, ETH3D, and DIODE, comparing against MiDaS v3.1's strongest model (DPT-BEiT-L). Using a DINOv2 encoder with DPT decoder performing depth regression, the teacher model is trained for 20 epochs on labeled images and the student model sweeps unlabeled images once. Results with the ViT-L encoder show dramatic improvements: KITTI AbsRel drops from 0.127 (MiDaS) to 0.076; NYUv2 from 0.048 to 0.043; Sintel from 0.587 to 0.458; DDAD from 0.251 to 0.230; ETH3D from 0.139 to 0.127; DIODE from 0.075 to 0.066. Remarkably, the ViT-B model (97.5M parameters) is already clearly superior to MiDaS based on the much larger ViT-L, and even the ViT-S model (24.8M parameters) outperforms MiDaS on several datasets despite using approximately 1/10 the parameters.

When fine-tuned to metric depth estimation, the model achieves new state-of-the-art on both indoor and outdoor benchmarks. On NYUv2: δ₁ reaches 0.984 (vs VPD's 0.964), AbsRel drops to 0.056, and RMSE to 0.206. On KITTI: δ₁ reaches 0.982 (vs NDDepth's 0.978), AbsRel at 0.046, RMSE at 1.896. Replacing MiDaS encoder with Depth Anything in the ZoeDepth framework for zero-shot metric depth also shows consistent improvements across unseen datasets. Furthermore, the pre-trained encoder demonstrates strong transferability to semantic segmentation: Cityscapes achieves 86.2 mIoU (surpassing Swin-L's 84.6 and Mask2Former's 84.3), and ADE20K reaches 59.4 mIoU (improving from previous best of 58.3). This supports the claim that Depth Anything has great potential to serve as a generic multi-task encoder.

Ablation studies systematically verify the contribution of each proposed component. Using only labeled data yields a mean AbsRel of 0.085. Simply adding unlabeled images with pseudo labels does not necessarily bring gains (mean AbsRel remains 0.085), confirming that naive self-training is insufficient when labeled data is already adequate. Adding strong perturbations reduces mean AbsRel to 0.081, and further incorporating the feature alignment loss L_feat achieves the best result of 0.076. Individual dataset contribution analysis reveals that HRWSI (only 20K images) provides the strongest generalization despite being the smallest labeled dataset, emphasizing the importance of data diversity over sheer volume. Comparison with original DINOv2 shows consistent improvements: NYUv2 AbsRel improves from 0.066 to 0.056; KITTI from 0.058 to 0.046; ADE20K mIoU from 58.8 to 59.4.

Further ablations examine the tolerance margin α in the feature alignment loss. Testing values of α = 0.00, 0.15, and 0.30 reveals that α = 0.15 is optimal (mean AbsRel 0.175), while α = 0 yields 0.188 and α = 0.30 yields 0.178. Without the margin, excessive alignment forces the depth features to mirror DINOv2's tendency to produce uniform features within objects, conflicting with the need for part-level depth discrimination. Additionally, applying L_feat to unlabeled data is beneficial (mean AbsRel 0.175), but applying it to labeled data proves harmful (0.179 vs 0.180 baseline). This is attributed to pseudo-labeled data's noisiness benefiting from semantic constraints, while manually labeled data's quality suffers from the interference of the alignment signal.

This work highlights the value of cheap, diverse unlabeled images for monocular depth estimation. Two simple yet highly effective strategies fully exploit their potential: (1) posing challenging optimization targets when learning from unlabeled images through strong perturbations, and (2) preserving semantic priors from pre-trained models via feature alignment. The resulting model demonstrates excellent zero-shot depth estimation ability and also serves as a promising initialization for downstream metric depth estimation and semantic segmentation tasks. Currently, the largest model size is constrained to ViT-Large. In the future, the authors plan to further scale up the model size from ViT-Large to ViT-Giant and increase training resolution from 512 to 700+ or 1000+ pixels for real-world applications.