Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting (CVPR'26 Findings)

Re-Depth refines the prediction of the monocular depth estimators via novel, self-supervised re-lighting method.

This is the official repository that contains source code for the arXiv paper Re-Depth Anything.

[Paper] [Project Page]

If you find Re-Depth Anything useful for your work please cite:

@article{bhattarai2025redepth,
      title={Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting},
      author={Ananta R. Bhattarai and Helge Rhodin},
      journal={ArXiv},
      year={2025},
      volume={abs/2512.17908}
}

News

• 15.02.2026: Code for Re-Depth Anything is released. 🔥

Prepare Environment

1. Clone the repository, then run the scripts to configure your environment and download required assets:

git clone https://github.com/anantarb/Re-Depth-Anything.git
cd Re-Depth-Anything/
source ./scripts/install_deps.sh
sh ./scripts/prepare_DAv2_small.sh

The first script creates the redepth conda environment and installs all necessary packages. The second script downloads the pre-trained models and supporting modules, and places them into the appropriate folders.

This project was developed and tested on Ubuntu 24.04, CUDA 12.8, and Python 3.12. If you are using a different OS, CUDA version, or Python version, you may need to make additional adjustments to get the environment working.

If you run into dependency issues, you can refer to requirements_dev.txt, which lists the exact package versions used during development.

Running

Preparing the config

We provide an example config at redepth/config/DAv2small_example.yaml. To run Re-Depth on your own image, make the following changes:

• image_path: Set this to the path of your input image.
• depth_path (optional): Set this to the path of your ground-truth (GT) depth map, if available. When provided, the output will be compared against GT both qualitatively and quantitatively, and the results will be saved in the logging directory.
• depth_min_value / depth_max_value (optional): Adjust these if you want to evaluate only GT depth values within a specific range. If not set, the defaults are:
  • depth_min_value = 1e-3
  • depth_max_value = 1e4
• mask (optional): Set this if you have an image mask. Note that passing mask does not affect the optimization process.
• depth_mask (optional): Set this if you want evaluation against GT to be limited to the regions defined by depth_mask. If depth_mask is not provided, it is automatically derived from the GT depth values using depth_min_value and depth_max_value.
• depth_to_meters (optional): Set this scaling factor to convert the GT depth units to meters for evaluation.
• save_dir: Set this to the directory where you want outputs to be saved.

For details on how inputs are loaded and processed, see redepth/dataset/DAv2_dataset.py.

The items above are the minimum changes needed to run Re-Depth on your own image. For best results, you may want to tune additional hyperparameters such as lr, scale, smoothness_weight, and max_steps.

Running Re-Depth Anything

After configuring your setup, run the optimization script:

python redepth/scripts/run_redepth_DAv2.py --config=PATH_TO_CONFIG

Output directory + naming conventions

Outputs are saved to save_dir every save_interval (as specified in your config). The value {X} (see below) is computed as: X = (current_ensemble_size * max_step + current_step)

where:

• current_ensemble_size ∈ [0, ensemble_size - 1]
• X = 0 always refers to the initial model output (no optimization applied).

Depth outputs (ensembled)

• 0_{X}.png: Depth visualized using the cv2.COLORMAP_INFERNO colormap.
• {X}_raw.pt: Raw model output with no transformations applied.
• {X}_scaled.pt: Scaled orthographic model output (Eq. 5 in the main paper).
• {X}.json: Quantitative evaluation of the output against GT depth.
• GT.pt: Resized ground truth depth.

Normal outputs (computed from ensembled depths)

• 0_{X}.png: Normals computed from the scaled orthographic depth using Eq. 3 in the main paper.

Model outputs (not ensembled)

• {X}.pt
  • Checkpoints are saved from the current step within each ensemble run (the models themselves are not ensembled).

Stacked outputs

• {X}.png: Input image, GT depth (if available), output depth (ensembled), and normals from the scaled orthographic output depth (ensembled) stacked together for visualization.

Applying Re-Depth to Other Models

• Re-Depth Anything can be used with other backbone models (e.g., Depth-Anything-V2-Giant, DA3MONO-LARGE). To do this, create a model-specific coach/trainer class (see redepth/coach/DAv2_coach.py) that inherits from the base coach at redepth/coach/base_coach.py.
• The base coach expects the derived class to provide the required attributes (e.g., Dataset, Model, Optimizer) and to implement the required methods (listed under @abstractmethod) so the optimization can be launched.
• Keep in mind that different base models output depth/disparity in different value ranges. As a result, the configuration values for scale and ms may need to be significantly different across models. Since Re-Depth relies on the scaled depth computed from these parameters, getting good estimates for scale and ms is important for best performance.
• Other hyperparameters can also strongly affect results, including smoothness_weight, embeddings_lr, dpt_lr, and scale_lr.
• We plan to expand the codebase to support more base models in the future.

Acknowledgements

Our work builds on top of amazing open-source networks and codebases. We thank the authors for providing them.

• Depth Anything V2: a SOTA monocular depth estimator.
• threestudio: a unified framework for 3D content creation from text prompts, single images, and few-shot images, by lifting 2D text-to-image generation models.
• Hugging Face: a platform that provides libraries for many machine learning tasks like text generation, image generation, and many more.