code

DADO Code Snapshot

This folder contains the DADO implementation used for the CAIP 2025 paper.

Contents

• dado.py: main DADO pipeline.
• myutils/: local DADO utility modules.
• dinov1/: local DINOv1 implementation used by the default configuration.
• dinov2/: included because the main script imports the DINOv2 loader, although the default configuration uses DINOv1.
• prepare_project.sh: setup script for dependencies, local folders, the DINOv1 checkpoint, and optional VOC datasets.
• requirements.txt: Python runtime dependencies.

Large local assets are intentionally not tracked in Git:

• pretrained/
• datasets/
• results/
• analysis/

Use the setup script below to recreate the required local files after cloning.

Prepare

Create and activate a Python environment first. For example, with venv:

python3.11 -m venv .venv
.venv/bin/python -m pip install --upgrade pip

Then prepare the project from this folder with the venv interpreter:

PYTHON=.venv/bin/python ./prepare_project.sh

Or, with conda:

conda create -n dado python=3.11
conda activate dado

Then prepare the project from this folder:

./prepare_project.sh

This installs requirements.txt, creates the local output/data folders, and downloads the required DINOv1 checkpoint for the default DADO configuration:

pretrained/dino_deitsmall16_pretrain.pth

The original project also kept other optional pretrained weights for experiments. To download the full local pretrained set, run:

./prepare_project.sh --with-all-pretrained

That downloads:

• pretrained/dino_deitsmall16_pretrain.pth default DINOv1 small checkpoint
• pretrained/dino_vitbase16_pretrain.pth optional DINOv1 base checkpoint
• pretrained/dinov2_vitb14_reg4_pretrain.pth optional DINOv2 base checkpoint with registers
• pretrained/dinov2_vits14_pretrain.pth optional DINOv2 small checkpoint without registers
• pretrained/dinov2_vits14_reg4_pretrain.pth optional DINOv2 small checkpoint with registers
• pretrained/depth_anything_v2_vitl.pth optional Depth Anything V2 large checkpoint

To also download Pascal VOC 2007:

./prepare_project.sh --with-voc2007

To download VOC 2007 and VOC 2012:

./prepare_project.sh --with-all-voc

If you already have an environment prepared and only want the checkpoint/folders:

./prepare_project.sh --no-deps

If your Python executable is not named python, pass it explicitly:

PYTHON=/path/to/python ./prepare_project.sh --with-voc2007

Run

From this folder:

MPLBACKEND=Agg python dado.py --dataset-path /path/to/VOC2007/VOCdevkit/VOC2007 --images 009331 --timestamp test_009331

If you downloaded VOC 2007 with prepare_project.sh --with-voc2007, you can run:

MPLBACKEND=Agg python dado.py --dataset-path datasets/VOC2007/VOCdevkit/VOC2007 --images 009331 --timestamp test_009331

If --images is omitted, DADO processes all images returned by the VOC image-set loader.

Outputs are written locally to:

• results/<timestamp>_results.csv
• results/<timestamp>_images/
• analysis/<timestamp>_images/

The DPT depth model is loaded from Hugging Face (Intel/dpt-beit-base-384) unless it is already present in the local Transformers cache.