Depth Estimators Are Implicit Neural Fields for 3D Scene Geometry Inpainting and Reconstruction

Code release for Depth Estimators Are Implicit Neural Fields for 3D Scene Geometry Inpainting and Reconstruction.

This repository contains the paper-release scripts for:

• Birmingham satellite RGB-D inpainting and InfiniDepth finetuning;
• evaluation helpers, manifests, and lightweight Birmingham result previews.

The Birmingham finetuned checkpoint is hosted on Hugging Face:

huggingface-cli download Behavia/NeuralDepthField \
  checkpoints/birmingham/d2_overlap_consistency_epoch015.ckpt \
  --local-dir checkpoints_and_results

Repository Layout

scripts/birmingham/       Birmingham stage-1/2/3 preprocessing, completion, and export scripts
scripts/report/           Figure/table helper scripts
infinidepth_patches/      Files to overlay onto the InfiniDepth codebase
docs/                     Experiment design, runtime notes, and paper-ready summaries
results/                  Small tables, summaries, and preview figures

Environment

The experiments use InfiniDepth plus the release overlay in this repository.

git clone https://github.com/Shadow-Dream/Neural-Depth-Field.git
cd Neural-Depth-Field

# Clone or place the InfiniDepth codebase next to this repo.
git clone <INFINIDEPTH_REPO_URL> third_party/InfiniDepth
rsync -a infinidepth_patches/ third_party/InfiniDepth/

conda create -n ndf python=3.12 -y
conda activate ndf

# Use a CUDA build that matches your machine. The Pro 6000 runs used cu128.
pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128
pip install -r third_party/InfiniDepth/requirements.txt
pip install -r requirements-release.txt

export INFINIDEPTH_ROOT=$PWD/third_party/InfiniDepth
export PYTHONPATH=$INFINIDEPTH_ROOT:$PWD

Validated training environments:

• Local smoke machine: Python /data/clawdex-cityrepair/envs/infinidepth-py310/bin/python, PyTorch 2.5.1+cu121, RTX 4090.
• Pro 6000 training machine: Python /data/data_jyz/uav/envs/infinidepth-py312-cu128/bin/python, PyTorch 2.11.0+cu128, RTX PRO 6000 Blackwell.

The Pro 6000 environment is the canonical environment for the Birmingham checkpoint.

Birmingham Result

The Birmingham release checkpoint is:

checkpoints/birmingham/d2_overlap_consistency_epoch015.ckpt

It is the InfiniDepth depth-sensor model finetuned on Birmingham satellite RGB-D windows with overlap consistency. The key Birmingham depth-completion results are height MAE in meters; lower is better.

Evaluation protocol	Baseline MAE	Released checkpoint MAE	Reduction
Structured missing-region masks	0.5842	0.3111	46.7%
Realistic invalid-mask transfer	0.3347	0.2226	33.5%
Paired-window overlap consistency	0.2398	0.0880	63.3%

The overlap-consistency continuation also improves the strongest finetuned baseline from 0.3208 to 0.3111 on structured masks, from 0.2249 to 0.2226 on realistic hole-transfer masks, and from 0.1085 to 0.0880 on the paired-window consistency benchmark.

Preview artifacts included in this repository:

• results/birmingham/01_global_raw_vs_completed.png
• results/birmingham/completed_height_preview_1_8.png
• results/birmingham/sample00_srcwin0327_r09216_11264_c08448_10496_original_raw_2048.png
• results/birmingham/sample00_srcwin0327_r09216_11264_c08448_10496_sliding_completed_2048.png

The dense full-resolution Birmingham RGB-D arrays are not committed to GitHub. The release scripts record the exact local output paths used to generate them.

Citation

This release is prepared for the paper draft above. Add the final BibTeX entry here after paper metadata is fixed.