HeightMatch

This repository provides the basic pipeline for HeightMatch using INRIA_Buildings as an example.

Workflow

The recommended order is:

1. Prepare the pretrained checkpoint
2. Generate synthetic height maps
3. Train the model
4. Test the model

---

0. Prepare the pretrained checkpoint

Before training, make sure the pretrained DINOv2 checkpoint is available in the pretrained folder.

Required file:

pretrained/dinov2_small.pth

Download link:

DINOv2-Small

After downloading, rename the file to:

dinov2_small.pth

and place it in:

pretrained/

---

1. Generate synthetic height maps

Run generate_synthetic_height.py to generate synthetic height maps from RGB images.

Example for one city:

python3 generate_synthetic_height.py \
    --imgdir /home/Datasets/RSSeg/INRIA/Austin/image \
    --outdir /home/Datasets/RSSeg/INRIA/Austin/height

Important note

The script only accepts one imgdir and one outdir at a time.
Therefore, if the INRIA dataset is organized by multiple cities, you need to run the script once for each city.

Example:

python3 generate_synthetic_height.py \
    --imgdir /home/Datasets/RSSeg/INRIA/Austin/image \
    --outdir /home/Datasets/RSSeg/INRIA/Austin/height

python3 generate_synthetic_height.py \
    --imgdir /home/Datasets/RSSeg/INRIA/Chicago/image \
    --outdir /home/Datasets/RSSeg/INRIA/Chicago/height

The generated height maps are saved in the folder specified by --outdir.

---

2. Prepare split files

Before training, make sure the split files are ready, for example:

splits/inria_buildings/
├── val.txt
├── test.txt
└── 0.5/
    ├── labeled.txt
    └── unlabeled.txt

Each line should contain:

image_path label_path height_path

Example:

Austin/image/austin1.tif Austin/label/austin1.tif Austin/height/austin1.png

The third path must point to the generated synthetic height map.

---

3. Train

Training is launched with train.sh.

Example settings in train.sh:

dataset='inria_buildings'
split='0.5'
method='heightmatch'
model='dinov2_small'

Run training:

bash train.sh 4

Here, 4 means using 4 GPUs.

The checkpoints are saved to:

saved/inria_buildings/0.5/heightmatch/dinov2_small/

---

4. Test

Testing is launched with test.sh.

Make sure the settings in test.sh are consistent with training:

dataset='inria_buildings'
split='0.5'
method='heightmatch'
model='dinov2_small'

Run testing:

bash test.sh

The testing script loads:

saved/inria_buildings/0.5/heightmatch/dinov2_small/best.pth

The results are written to:

test_results.xlsx

---

5. Minimal example

# Step 0: prepare pretrained checkpoint
# Put pretrained/dinov2_small.pth in the pretrained folder

# Step 1: generate synthetic height maps
python3 generate_synthetic_height.py \
    --imgdir /home/Datasets/RSSeg/INRIA/Austin/image \
    --outdir /home/Datasets/RSSeg/INRIA/Austin/height

# Repeat for other cities if needed

# Step 2: train
bash train.sh 4

# Step 3: test
bash test.sh