Depth Benchmark

Benchmark monocular depth estimation and segmentation models on aerial imagery (SkyScenes dataset).

Installation

conda env create -f environment.yml
conda activate depth_benchmark
pip install -e .

# For segmentation (optional)
pip install transformers

Download SkyScenes

# Download images + depth
bash scripts/download_skyscenes.sh --path /path/to/SkyScenes

# Download segmentation masks (ClearNoon only)
bash scripts/download_skyscenes_segmentation.sh --path /path/to/SkyScenes

# Extract all archives (images, depth, segmentation)
bash scripts/extract_skyscenes.sh /path/to/SkyScenes

Depth Benchmarking

python scripts/evaluate_skyscenes.py \
    --dataset /path/to/SkyScenes \
    --model moge \
    --output results/

# Filter by altitude/pitch/weather/town
python scripts/evaluate_skyscenes.py \
    --dataset /path/to/SkyScenes \
    --model moge \
    --altitudes 15 30 \
    --pitches 0 -45 \
    --weathers ClearNoon \
    --towns Town01 Town02

Segmentation Benchmarking

# SegFormer (neural network)
python scripts/evaluate_skyscenes_segmentation.py \
    --dataset /path/to/SkyScenes \
    --model segformer-b5-cityscapes \
    --output results/

# RANSAC (geometry-based, uses GT depth)
python scripts/evaluate_skyscenes_segmentation.py \
    --dataset /path/to/SkyScenes \
    --model ransac \
    --output results/

# RANSAC with predicted depth from MoGe
python scripts/evaluate_skyscenes_segmentation.py \
    --dataset /path/to/SkyScenes \
    --model ransac \
    --depth-model moge \
    --output results/

# Filter by altitude/pitch/town
python scripts/evaluate_skyscenes_segmentation.py \
    --dataset /path/to/SkyScenes \
    --model segformer-b5-cityscapes \
    --altitudes 15 \
    --pitches 0 \
    --towns Town01

Available Models

Depth: moge, moge-2-vitl

Segmentation: segformer-b0-cityscapes, segformer-b5-cityscapes, ransac