# NvDepthAnythingV2 Overview ### Description: DepthAnythingV2 is a state-of-the-art monocular depth estimation model that generalizes well on zero-shot images. TAO NvDepthAnythingV2 relative depth model is a pretrained commercial model that can produce a relative depth map given a single image. In addition, NvDepthAnythingV2 can be used as initialization when fine-tuning the MetricDepthAnything model for better accuracy.
This model is ready for commercial use.
### License/Terms of Use License to use these models is covered by the NVIDIA Open Model License. By downloading the model, you accept the terms and conditions of these [licenses](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). ### Deployment Geography: Global
### Use Case:
This model is intended for developers working on industrial, robotics, and smart space applications to estimate the depth from monocular image input.
### Release Date:
NGC [07/25/2025] [link](https://registry.ngc.nvidia.com/orgs/nvstaging/teams/tao/models/nvdepthanythingv2)
## References(s): [DepthAnythingV2 paper](https://arxiv.org/pdf/2406.09414)
## Model Architecture: **Architecture Type:** Transformer-based Network Architecture
**Network Architecture:** Vit-Large
**Number of model parameters:** 3.6*10^8 **This model was developed based on [DINOV2-ViT-Large](https://github.com/facebookresearch/dinov2/tree/main)
## Computational Load **Cumulative Compute:** 2.0952*10^16
**Estimated Energy and Emissions for Model Training:** * Energy Consumption in kWh: 1075.2 kWh * Emissions: 0.3472896 tCO2e ## Input:
**Input Type(s):** RGB image
**Input Format:** Red, Green, Blue (RGB)
**Input Parameters:** Two-Dimensional (2D)
**Other Properties Related to Input:** B X 3 X H X W (Batch Size x Channel x Height x Width)
## Output:
**Output Type(s):** Image
**Output Format:** Depth Map Image
**Output Parameters:** Two-Dimensional (2D)
**Other Properties Related to Output:** B X 3 X H X W (Batch Size x Channel x Height x Width)
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
## Software Integration: **Runtime Engines** * TAO 6.2.0
**Supported Hardware Microarchitecture Compatibility:**
* NVIDIA Ampere
* NVIDIA Blackwell
* NVIDIA Jetson
* NVIDIA Hopper
* NVIDIA Lovelace
* NVIDIA Pascal
* NVIDIA Turing
* NVIDIA Volta
**Preferred/Supported Operating System(s):** * Linux
* Linux 4 Tegra
* QNX
* Windows
## Model Version(s): - **deployable_relative_depthanythingv2_large_v1.0**: decrypted [ONNX](https://onnx.ai/) files. Inference supported in [TAO Toolkit](https://github.com/NVIDIA/tao_pytorch_backend).
## Training, Testing, and Evaluation Datasets: ** The total size (in number of data points): 6.02M images
** Total number of datasets: 4 datasets
** Dataset partition: Training 5.99M, validation 29K
### Training Dataset: Data Collection Method by dataset:
* Hybrid: Automatic/Sensors, Synthetic
Labeling Method by dataset:
* Hybrid: Automatic/Sensors, Synthetic
**Properties:** | Dataset | No. of Images |Sensor| |--|--|--| |NV Internal Real Image Data| 3.6M | Automated | |NV Internal Real Image Data| 94K | RGB Camera| |NV Internal Synthetic Data | 2M | Synthetic | |Crestereo Synthetic Data | 196K | Synthetic |
### Evaluation Dataset: **Link:** - NYUDV2 [Dataset Page](https://cs.nyu.edu/~fergus/datasets/nyu_depth_v2.html)
**Benchmark Score
Accuracy was determined using the following metrics: - AbsRel : (Scale-Shift Invariant) Absolute Relative Error - D-1 : (Scale-Shift Invariant) This measures the percentage of predicted pixels that differ from the true pixels by no more than 25%. The results is zero-shot evaluation on the NYUDV2 dataset and relative depth (inverse depth) with scale-shift invariant accuracy is used to calculate the metric. | Method | AbsRel | D-1 | | ------- | ------- | -------| | NvDepthAnythingV2 | 0.047 | 0.979 | Data Collection Method by dataset:
* Undisclosed
Labeling Method by dataset:
* Automatic/Sensors
**Properties:** 654 images and ground truth depth were used for evaluation.
# Inference **Acceleration Engine:** Onnx
**Test Hardware [Name the specific test hardware model]** - Jetson AGX Orin - Jetson AGX Thor - L4 - L40S - A100 - RTX PRO 6000 Blackwell - H200 - H100 - B200 - GB200 The inference performance of the provided NvRelativeDepthAnything model is evaluated at FP16 precisions. The model's input resolution is 3x518x914 pixels. The performance assessment was conducted using trtexec on a range of devices.
| Platform | BS | FPS | |------------------|----|------| | Jetson AGX Orin | 32 | 3.705279734 | | Jetson AGX Thor | 32 | 3.338531822 | | L4 | 32 | 17.89128806 | | L40S | 32 | 62.5196595 | | A100 | 32 | 73.9937429 | | RTX PRO 6000 Blackwell | 32 | 112.9716124 | | H200 | 32 | 166.6069658 | | H100 | 32 | 172.7824453 | | B200 | 32 | 320.4046711 | | GB200 | 32 | 350.4049915 | ## Ethical Considerations: NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards. Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).