SEEDepth

This work presents Semantic-Enhanced and Efficient Distillation for Depth Estimation (SEEDepth), a two-stage framework that effectively fuses information of depth and semantic foundation models to enhance Monocular Depth Estimation (MDE).

Usage

Installation (Conda environment)

• Python 3.11.7

git clone https://github.com/wodon326/SEEDepth.git
cd SEEDepth
pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Models

We provide two models of varying scales for relative depth estimation based on DepthAnything-V1 and DepthAnything-V2:

Model	# Params
DepthAnything-V1-Base	98M
DepthAnything-V1-Base + SEEDepth	105M
DepthAnything-V1-Large	335M
DepthAnything-V1-Large + SEEDepth	349M
DepthAnything-V2-Base	98M
DepthAnything-V2-Base + SEEDepth	105M
DepthAnything-V2-Large	335M
DepthAnything-V2-Large + SEEDepth	349M

To run the model, you need to download each model above.

Run

python inference.py \
    --input_dir <input-dir> \
    --output_dir <output-dir> \
    --checkpoints_dir <checkpoints-dir> \
    --model_version <v1 | v2> \
    --model_size <base | large>

Arguments:

• --input_dir: Path to the folder containing input images (default: ./input_images)
• --output_dir: Path to the folder where prediction of disparity will be saved (default: ./output)
• --checktpoints_dir: Path to the directory where model checkpoints are stored (default: ./checkpoints)
• --model_version: Version of the model to use (default: v2)
• --model_size: Size of the model to use (default: large)

Acknowledgement

Thanks to these amazing open source projects:

• DepthAnything-V1
• DepthAnything-V2
• SegmentAnything
• Marigold
• Sample images in input_images are sampled from DIODE dataset.

Citation

Paper

@INPROCEEDINGS{11385908,
  author={Ma, Sanggyun and Choi, Wonjoon and Park, Jihun and Kim, Jaeyeul and Im, Sunghoon},
  booktitle={2026 International Conference on Electronics, Information, and Communication (ICEIC)}, 
  title={Semantic-Enhanced Monocular Depth Estimation via Fusion and Distillation of Foundation Models}, 
  year={2026},
  volume={},
  number={},
  pages={1-5},
  keywords={Training;Attention mechanisms;Foundation models;Fuses;Depth measurement;Computational modeling;Semantics;Benchmark testing;Transformers;Computational efficiency;Foundation model;Monocular depth estimation;Zero-shot depth estimation;Knowledge distillation},
  doi={10.1109/ICEIC69189.2026.11385908}
}