[IROS2025]BoRe-Depth

A novel lightweight monocular depth estimation method.

It is suitable for all kinds of real-time tasks, especially on embedded devices of unmanned systems. video

Performance

We provide two models (NYUv2 and KITTI) for robust relative depth estimation.

Dataset	Param./M	Abs_Rel	RMSE	$\delta$1	$\delta$2	$\delta$3	$\epsilon$DBEacc
NYUv2	8.7	0.101	0.429	0.883	0.971	0.993	2.083
KITTI	8.7	0.103	4.323	0.889	0.967	0.986	2.649

Indoor Scene Visualization

Outdoor Scene Visualization

Usage

Installation

git clone https://github.com/liangxiansheng093/BoRe-Depth.git
cd BoRe-Depth
pip install -r requirements.txt

Download weights to checkpoints folder.

Test

python test.py --dataset_name nyu --dataset_dir datasets/nyu/testing --ckpt_path checkpoints/nyu.ckpt

Options:

• --dataset_name: [nyu, kiit, iBims]. The size of the predicted depth map based on the selected dataset.
• --dataset_dir: The path to the test dataset (both jpg and png format).
• --ckpt_path: The path to the trained weights.

Inference

python infer.py --dataset_name nyu --ckpt_path checkpoints/nyu.ckpt --input_dir demo --output_dir output --save-vis --save-depth

Options:

• --dataset_name: [nyu, kiit, iBims]. The size of the predicted depth map based on the selected dataset.
• --ckpt_path: The path to the trained weights.
• --input_dir: The path to the input picture or folder (both jpg and png format).
• --output_dir: The path to the output depth map (both png and npy format).
• --save-vis: Saving the visual images.
• --save-depth: Saving the numpy results.