AerialMetric: Benchmarking and Adapting UAV Monocular Metric Depth Estimation in the Real World

Zhongqiang Song¹    Guanying Chen^1,✉    Yuqi Zhang^2,3    Yin Zou¹

Chuanyu Fu¹    Zhiyuan Yuan¹    Chuan Huang^4,2    Shuguang Cui^3,2    Xiaochun Cao¹

¹ Sun Yat-sen University, Shenzhen Campus      ² FNii-Shenzhen      ³ SSE, CUHKSZ      ⁴ SIAS, USTC

AerialMetric is a real-world UAV monocular metric depth benchmark covering oblique photogrammetry, controlled aerial variables, in-the-wild UAV videos, and adapted MoGe-2 baselines.

News

• ✅ 2026-07-24: The code repository is now publicly available and includes both inference and benchmark scripts.
• 🎉 2026-06-18: AerialMetric was accepted to ECCV 2026.

Benchmark Entry Points

cd /path/to/AerialMetric

The repository is organized into three benchmark interfaces:

Directory	Purpose
MoGe/	Aerial benchmark for MoGe2 and AerialMetric
Ground_MoGe/	Ground benchmark for MoGe2 and AerialMetric
Depth_Baselines/	Ground benchmark wrappers for UniDepth, DepthPro, and ZoeDepth

Dataset layout requirements are documented in DATA_ORGANIZATION.md.

Datasets & Weights

Pipeline of AerialMetric.	Dataset composition of AerialMetric.

Resource	Link
Dataset & Weights	Kuiee/AerialMetric-ECCV2026

Environment Setup

conda create -n mogefresh3 python=3.10 -y
conda activate mogefresh3
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
pip install -e MoGe/

Verify:

python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA: {torch.cuda.is_available()}')"
python -c "from moge.model.v2 import MoGeModel; print('MoGe OK')"

Reproduction

Edit paths in benchmark.sh to match your local setup, then run:

# 4 GPUs in parallel (recommended)
bash benchmark.sh 0,1,2,3

# Single GPU
bash benchmark.sh 0

The script runs all experiments automatically:

Type	Experiment	Datasets	Intrinsics	Mask	Models
Aerial Scenes	1	Decoupled + Oblique + Wild	none	load	Baseline + MoGe2-Aerial LoRA
Aerial Scenes	2	Decoupled + Oblique (norm)	load	load	Baseline + MoGe2-Aerial LoRA
Ground Scenes	3	Ground ×7 (NYUv2, KITTI, …)	no oracle	—	Baseline + MoGe2-Aerial LoRA
Ground Scenes	4	Ground ×7	oracle (GT intrinsics)	—	Baseline + MoGe2-Aerial LoRA

Results are saved to OUTPUT_ROOT (default: /data1/szq/622-freash3) with per-dataset evaluation reports (.txt / .json).

Weights:

• MoGe2(Baseline): MoGe-2 ViT-Large
• MoG2-Aerial(LoRA Type): AerialMetric-ECCV2026

Dataset layout: see DATA_ORGANIZATION.md. Ground config: edit Ground_MoGe/configs/eval/ground_metric_benchmarks_local.json.

Quick Demo — Single Image / Video / Folder Inference

# Base model (single image)
python demo_infer.py \
  --input photo.jpg \
  --model vitl \
  --checkpoint /path/to/vitl-normal.pt \
  --output ./demo_output

# Aerial LoRA model (video, every 5th frame)
python demo_infer.py \
  --input video.mp4 \
  --model lora \
  --checkpoint /path/to/Moge2-Aerial.pt \
  --lora_config MoGe/configs/Final_train/config-lora-all.json \
  --output ./demo_output \
  --stride 5

# Full options example
python demo_infer.py \
  --input ./images/ \
  --model lora \
  --checkpoint /path/to/Moge2-Aerial.pt \
  --lora_config MoGe/configs/Final_train/config-lora-all.json \
  --output ./demo_output \
  --resize 1024 \
  --resolution_level 9 \
  --cmap inferno \
  --save_components

CLI Reference

Category	Argument	Default	Description
Required	--input	—	Image / video / folder path
	--model	—	vitl (base) or lora (LoRA fine-tuned)
	--checkpoint	—	Path to .pt checkpoint
LoRA	--lora_config	—	LoRA config JSON (required for --model lora)
	--lora_rank	96	LoRA rank
Inference	--resize	0	Long-edge resize target (0=original). Padded to ×14
	--resolution_level	9	MoGe2 quality/speed: 0 (fastest) ~ 9 (best)
	--force_projection	on	Recompute point map from depth for consistency
	--apply_mask	on	Mask invalid depth regions
	--fov_x	auto	Override horizontal FoV (degrees), auto-estimate if omitted
	--device	cuda	cuda / cpu
	--fp16 / --no_fp16	on	FP16 inference
Output	--output	./demo_output	Output directory
	--save_npy / --no_npy	on	Save raw .npy depth array
	--save_components	off	Also save standalone _depth_vis.png
Visualization	--cmap	jet	Colormap: jet / inferno / plasma / viridis / turbo
	--vmin / --vmax	auto	Depth range clamp for colormap
	--seed	42	Random seed for point picking
	--point_margin	20	Edge margin for point sampling
Video	--stride	1	Process every N-th frame

Output per frame

File	Description
*_depth.npy	Raw float32 depth array
*_depth_annotated.png	Composite: depth colormap + 2 random point depth markers + colorbar
*_colorbar.png	Standalone depth scale bar
*_meta.json	Per-frame metadata (see below)

Weights:

• MoGe2(Baseline): MoGe-2 ViT-Large
• MoG2-Aerial(LoRA Type): AerialMetric-ECCV2026

Acknowledgments

Models & Tools

• MoGe-2 (Wang et al., 2025) — the core monocular geometry estimation model used in this work.
• Gemini and Depth Anything 3 — used to assist in the construction and processing of the AerialMetric-Wild subset.
• DJI Terra — the primary software used for aerial photogrammetric reconstruction and geometric data processing.
• UniDepthV2 (Piccinelli et al., 2025) — used to evaluate the cross-architecture generalization of models adapted using AerialMetric.

Oblique：Data Sources

We gratefully acknowledge the following datasets or works for providing raw UAV imagery, scene assets

• GauU-Scene — large-scale UAV aerial scene dataset.
• UrbanBIS (Yang et al., SIGGRAPH 2023) — benchmark for fine-grained urban building instance segmentation.
• UrbanScene3D (Lin et al., ECCV 2022) — large-scale urban scene dataset with high-resolution aerial images.
• UAVScenes (Wang et al., ICCV 2025) — multi-modal UAV dataset with frame-wise semantic annotations.
• OpenDroneMap — open-source toolkit for generating maps, point clouds, 3D models, and DEMs from drone images.
• ArcGIS Drone2Map (ESRI) — create 3D products from drone imagery.

Citation

If you find this work useful, please cite:

@inproceedings{song2026aerialmetric,
  title     = {AerialMetric: Benchmarking and Adapting UAV Monocular Metric Depth Estimation in the Real World},
  author    = {Zhongqiang Song and Guanying Chen and Yuqi Zhang and Yin Zou and Chuanyu Fu and Zhiyuan Yuan and Chuan Huang and Shuguang Cui and Xiaochun Cao},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Website License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.