Test-Time Prompt Tuning for Zero-Shot Depth Completion

Chanhwi Jeong · Inhwan Bae · Jin-Hwi Park* · Hae-Gon Jeon*
ICCV 2025 Highlight Paper
_{* Corresponding Authors}

📄 ICCV 2025 Paper 🌐 Project Page 💻 Official Code 🧩 Poster

Summary:

TestPromptDC introduces a test-time prompt tuning framework for zero-shot depth completion across varying sensors and environments, allowing foundation models for monocular depth estimation to achieve absolute-scale predictions without ground-truth fine-tuning.

Dataset Download Instructions

1. IBIMS-1

• Visit IBIMS-1 Website to download the dataset.
• After downloading, your folder structure should look like:

2. DDAD

• Follow the instructions at this repository to download the DDAD dataset.
• Once downloaded, place the files according to the following structure:

Dataset Path Setup

• For DDAD testing, set the --dataset_path argument to the location of ddad_train_val or the folder containing the ddad structure above.
• For IBIMS-1, set the --dataset_path argument to the ibims1 folder.

Usage

Run the following command, replacing each bracketed item with the appropriate value:

• --gpu: Specify the GPU number (e.g., 0 for the first GPU).
• --mode: Select either VP (visual prompt mode) or FT (fine-tuning mode).
• --dataset: Choose ibims or ddad.
• --dataset_path: Provide the path to your dataset folder.

Example: python main.py --gpu 0 --mode VP --dataset ddad --dataset_path /path/to/ddad

Additional Datasets (coming soon)

We will soon update download and configuration instructions for the following datasets:

• NYU Depth V2
• VOID
• KITTI Depth Completion
• nuScenes Depth

Quick Experiments: If you want to quickly test our framework or explore related setups, check out the following repositories:

🔹 DepthPrompting — Visual prompt tuning for depth-aware foundation models

🔹 UniDC — Universal Depth Completion baseline across sensors and environments

@InProceedings{Jeong_2025_ICCV,
  author    = {Jeong, Chanhwi and Bae, Inhwan and Park, Jin-Hwi and Jeon, Hae-Gon},
  title     = {Test-Time Prompt Tuning for Zero-Shot Depth Completion},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month     = {October},
  year      = {2025},
  pages     = {9443--9454}
}