A Simple yet Effective Test-Time Adaptation for Zero-Shot Monocular Metric Depth Estimation (IROS 2025)

Rémi Marsal · Alexandre Chapoutot · Philippe Xu · David Filliat

U2IS - ENSTA-Paris

[Paper] $~$ [Video]

This repository contains the official implementation of the paper A Simple yet Effective Test-Time Adaptation for Zero-Shot Monocular Metric Depth Estimation which was accepted to IROS 2025.

Installation

Download the datasets from official websites. Follow intructions here to install the dgp library.

git clone https://gitlab.ensta.fr/ssh/monocular-depth-rescaling
cd monocular-depth-rescaling
pip install -r requirements.txt
cd metric_depth

Follow instructions to install OmniGlue.

Then, to use rescaling with structure-from-motion and OmniGlue, the matching keypoints must be extracted beforehand using:

python extract_omniglue_matches.py -d <kitti | ddad> --save_path /path/to/save/keypoints

Evaluation

Download Depth Anything V1 checkpoints without fine-tuning here.

python evaluate.py \
-m zoedepth \
--pretrained_resource="local::/path/to/the/checkpoint.pth" \
-d <kitti | nyu | ddad | diode_outdoor | diode_indoor | sunrgbd | ibims> \
-r <lidar | stereo | sfm> \

Additional arguments:

• Use --num_beams with -r lidar to specify the number of beams of the LiDAR.
• Use --matching_with <sift | omniglue> with -r sfm to select the keypoints extraction and matching approach.
• Use -t with -r sfm to specify the minimum pose translation to consider.
• Use --path_to_keypoints /path/to/OmniGlue/keypoints with --matching_with omniglue to load OmniGlue keypoints.

Acknowledgement

We would like to thanks the authors of Depth Anything for their great work and for sharing their code.