microsplat_benchmarking / SETUP_NOTES.md
namor2023's picture
Add files using upload-large-folder tool
166007a verified
|
Raw History Blame Contribute Delete
18.8 kB

Benchmarking Setup Notes

Three deblurring-focused 3D Gaussian Splatting papers installed and smoke-tested (train -> render -> eval) on real Deblur-NeRF data. Environment: RTX A4000 (sm_86, 16GB), host driver CUDA 13.2, host nvcc 12.0, gcc-13 default (gcc-12 also installed).

Network constraint affecting all three: repo.anaconda.com and conda.anaconda.org both time out β€” every repo's documented conda install path is dead here. pypi.org, download.pytorch.org, github.com, gitlab.inria.fr, drive.google.com all work. Each repo below used a different conda-free workaround.

Shared dataset: official Deblur-NeRF archive (2.81GB, all four subsets β€” real/synthetic x defocus/motion), cached at shared_data/deblur_dataset.zip and partially extracted at shared_data/real_defocus_blur/ and shared_data/real_camera_motion_blur/. Source: the Drive file id from the Deblurring-3DGS README works; the Drive folder id linked in the CoCoGaussian/BAGS READMEs 404s.

gdown --no-cookies --continue "https://drive.google.com/uc?id=1N24ipi52XXiPfqWWAxYeyaItNXrL7nSe" -O deblur_dataset.zip

A HuggingFace mirror (zhaoshiwen/deblurnerf-motion10-processed-data) also exists but is not interchangeable with the official zip β€” it's been re-undistorted, downscaled, and re-solved (different calibration, confirmed by comparing blurcoffee intrinsics across sources). Use the official zip for any numbers meant to be comparable across repos or to the papers.

Cross-repo dataset gotcha: don't pass -i images_4. Scenes ship both images/ (.jpg, full-res) and images_4/ (.png, quarter-res), but COLMAP images.bin records .jpg filenames, and each repo's loader fallback only rewrites .png->.jpg, never the reverse β€” so -i images_4 raises FileNotFoundError. Use -i images -r 4 instead (same resolution, no dataset edits).

Cross-repo dataset gotcha: blurcoffee (in real_camera_motion_blur) is the one scene, of 30, that isn't SIMPLE_PINHOLE β€” it's SIMPLE_RADIAL with a small but genuinely nonzero k1 = 0.001927328970778844 (~2px shift at the full-res corner, ~0.5px at -r 4). BAGS hard-asserts on any non-pinhole model with no fallback; see the BAGS section below for the undistortion fix (tools/undistort_simple_radial.py) used instead of a source patch. Once undistorted and rewritten as PINHOLE, the scene is valid for both CoCoGaussian and BAGS β€” use the shared undistorted copy (BAGS/data/blurcoffee_undist/, or the same output re-staged per-repo) rather than each repo's raw SIMPLE_RADIAL copy:

  • Raw SIMPLE_RADIAL copy: BAGS asserts and fails outright. CoCoGaussian trains via a poses_bounds.npy fallback but then reads focal from poses_bounds (1674.40) instead of cameras.bin (1676.45) and ignores the distortion β€” don't quote that PSNR as a real/comparable result.
  • Undistorted PINHOLE copy: both repos read the identical rewritten cameras.bin (focal 1676.4482617963429) with no source patch to either. A controlled 600-iteration A/B on CoCoGaussian (matched seed) showed the undistorted copy scoring +0.68 dB test PSNR over the raw radial copy (26.257 vs 25.576), consistent with the correction improving multi-view geometric consistency rather than confounding results.

Iteration counts must still match for any cross-repo comparison β€” the BAGS blurcoffee_undist number below (27.044 at 6100 iters) is not comparable to CoCoGaussian's 600-iteration smoke numbers.

Benchmark reproduction summary

All three repos were also run to their own official iteration count (not the short smoke test) and compared against their respective paper's published numbers, on the closest matching scene. No additional fixes were needed beyond what's already documented per-repo below β€” same install, same source patches (or lack thereof), same dataset. Just the smoke-test flag overrides were dropped in favor of each repo's official config.

Method Scene Iters Paper PSNR/SSIM/LPIPS Ours Ξ”PSNR
CoCoGaussian Caps 30000 24.62 / 0.7472 / 0.1334 24.49 / 0.7406 / 0.1386 βˆ’0.13
BAGS Caps 12000 24.16 / 0.725 / 0.159 24.26 / 0.727 / 0.153 +0.10
BAGS Coffee 12000 31.59 / 0.908 / 0.096 30.63 / 0.903 / 0.117 βˆ’0.96
Deblurring-3DGS Ball 20000 28.27 / 0.8233 / 0.1413 28.31 / 0.8267 / 0.1492 +0.04

All four are within ~1 dB of their paper, confirming each implementation reproduces its published results. BAGS/Coffee is the outlier, plausibly because our run used the properly-undistorted blurcoffee (correcting a small but real k1β‰ˆ0.0019 COLMAP distortion β€” see the blurcoffee fix below) while the paper's original run likely used the scene as shipped, distorted.


CoCoGaussian (CVPR 2025)

CoCoGaussian/ β€” no source patches. Python 3.10 venv via uv (no Docker needed).

Install. Repo ships no .gitmodules, so --recursive leaves submodules/{diff-gaussian-rasterization,simple-knn} empty. The gitlinks pin stock upstream 3DGS commits, so clone those explicitly:

git clone --recursive https://github.com/Jho-Yonsei/CoCoGaussian.git CoCoGaussian
cd CoCoGaussian/submodules && rmdir diff-gaussian-rasterization simple-knn
git clone --recursive https://github.com/graphdeco-inria/diff-gaussian-rasterization.git
git clone https://gitlab.inria.fr/bkerbl/simple-knn.git
(cd diff-gaussian-rasterization && git checkout 59f5f77e3ddbac3ed9db93ec2cfe99ed6c5d121d && git submodule update --init --recursive)
(cd simple-knn && git checkout 86710c2d4b46680c02301765dd79e465819c8f19)
cd ..

uv venv --python 3.10 _venv
VIRTUAL_ENV=$PWD/_venv uv pip install --python $PWD/_venv/bin/python \
  torch==2.4.1 torchvision==0.19.1 --index-url https://download.pytorch.org/whl/cu121
VIRTUAL_ENV=$PWD/_venv uv pip install --python $PWD/_venv/bin/python \
  "numpy<2" "scikit-image==0.19.3" configargparse plyfile tqdm imageio imageio-ffmpeg \
  opencv-python-headless lpips tensorboard matplotlib gdown

export CC=/usr/bin/gcc-12 CXX=/usr/bin/g++-12 TORCH_CUDA_ARCH_LIST="8.6" MAX_JOBS=8
VIRTUAL_ENV=$PWD/_venv uv pip install --python $PWD/_venv/bin/python --no-build-isolation \
  ./submodules/simple-knn ./submodules/diff-gaussian-rasterization

scikit-image==0.19.3 (which forces Python 3.10) is load-bearing: metrics.py:85 calls structural_similarity(..., multichannel=True), a kwarg removed in 0.20+ (silently swallowed, then raises win_size exceeds image extent). gcc-12 is required β€” nvcc 12.0 rejects the default gcc-13.

Dataset. defocuscaps (34 imgs, hold=5) from the shared archive.

Gotchas.

  • metrics.py's __main__ is dead code (NameError on unimported ArgumentParser / undefined evaluate()). render.py is the actual eval path β€” it prints PSNR/SSIM/LPIPS itself.
  • No configs/ directory ships despite a --config flag existing β€” pass all args on the CLI.
  • Defaults (start_blurring=1500, start_weight=1500, pts_iter=2500, densify_from_iter=500) are tuned for 30k-iteration runs; a short smoke test must scale them down or it silently degrades to vanilla 3DGS (no CoCo blur kernel, no weight compositing, no point densification exercised).
  • Don't relocate a dataset after training β€” cfg_args stores an absolute source_path that render.py reads back.

Reproduce (smoke test, 600/30000 iters):

./_venv/bin/python train.py -s data/defocuscaps -i images -r 4 --expname coco_smoke \
  --iterations 600 --start_blurring 100 --start_weight 200 --pts_iter 300 \
  --densify_from_iter 200 --test_iterations 600 --save_iterations 600 --checkpoint_iterations 600
./_venv/bin/python render.py -m output/coco_smoke --iteration 600

Result: test PSNR 21.78 / SSIM 0.5596 / LPIPS 0.4924 (19,765 gaussians, ~2 min, 4.4GB VRAM). render.py reproduced these exactly from the saved checkpoint. Full run: --iterations 30000 with all defaults restored.

Reproduce (full run, 30000 iters, official/default config):

./_venv/bin/python train.py -s data/defocuscaps -i images -r 4 --expname coco_full30k --eval

No flag overrides beyond dataset path/resolution β€” all thresholds (start_blurring, start_weight, pts_iter, densify_from_iter) at their code defaults. Runtime ~1.5h on the RTX A4000: throughput drops from ~80 it/s early on to ~8 it/s once the CoC multi-Gaussian rendering kicks in around iteration 1500-2000 (each iteration renders the base Gaussians plus M CoC Gaussian sets β€” 5 renders/iteration by default). No extra fixes needed beyond what's documented above. Result: test PSNR 24.49 / SSIM 0.7406 / LPIPS 0.1386 (training-time eval, [ITER 30000] in log.txt), vs. paper's 24.62 / 0.7472 / 0.1334 for Caps (arXiv 2412.16028, Table 8) β€” within noise.


BAGS (ECCV 2024)

BAGS/ β€” no source patches to BAGS itself (one added standalone tool). Docker (docker/Dockerfile, docker/Dockerfile.ext, docker/run.sh).

Install. BAGS pins torch 1.12.1+cu113, too far from host nvcc 12.0 for a plain venv, so this repo uses Docker instead:

git clone --recursive https://github.com/snldmt/BAGS.git
cd BAGS
cp requirements.txt docker/requirements.txt
sudo docker build -t bags:cu113 -f docker/Dockerfile docker/    # py3.8 + torch 1.12.1+cu113
sudo docker build -t bags:ready  -f docker/Dockerfile.ext .     # + diff-gaussian-rasterization, simple-knn @ sm_86
./docker/run.sh python train.py ...   # mounts repo at /workspace, runs as host uid:gid

Base image is nvidia/cuda:11.3.1-devel-ubuntu20.04, not pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel (the latter ships Python 3.7, which has no wheels for numpy 1.24 / open3d 0.18 / scikit-image 0.19).

If pip fails inside the container with CERTIFICATE_VERIFY_FAILED on some hosts but not others, a corporate TLS-intercepting root CA is trusted by the host but not the container β€” copy the host's CA certs into docker/certs/ and set PIP_CERT/REQUESTS_CA_BUNDLE/SSL_CERT_FILE/CURL_CA_BUNDLE (also needed at runtime: metrics.py downloads a 233MB VGG16 for LPIPS).

Upstream ships a stale prebuilt submodules/diff-gaussian-rasterization/build/

  • .egg-info β€” move it aside to force a clean sm_86 compile.

Dataset. defocuscaps (34 imgs, hold=5) β€” the scene in BAGS' own README example.

blurcoffee fix (no source patch β€” data fix instead). BAGS hard-asserts on any COLMAP camera model other than SIMPLE_PINHOLE/PINHOLE/OPENCV (scene/dataset_readers.py:95, no poses_bounds fallback), and blurcoffee is SIMPLE_RADIAL with real (if small) distortion. Rather than patch the assert β€” which would need to correctly special-case k1β‰ˆ0 and still risks silently discarding distortion β€” the scene was properly undistorted instead, added as tools/undistort_simple_radial.py:

K    = [[f,0,cx],[0,f,cy],[0,0,1]]     # f=1676.4482617963429
dist = [k1, 0, 0, 0]                    # k1=0.001927328970778844
cv2.undistort(img, K, dist, None, K)    # newCameraMatrix=K keeps intrinsics/FoV unchanged
# then rewrite sparse/0/cameras.bin as PINHOLE (f, f, cx, cy); images.bin/points3D.bin copied verbatim

COLMAP's SIMPLE_RADIAL term matches OpenCV's k1-only radial term exactly. Verified genuinely radial by checking mean pixel diff rises monotonically with distance from center (0.19 at center -> 2.49 at the edge, matching the predicted ~2px corner shift). BAGS then runs on the undistorted copy with no code changes:

./docker/run.sh python tools/undistort_simple_radial.py data/blurcoffee_radial_src data/blurcoffee_undist

Other gotchas.

  • --ms_steps can't go below 6000 (assert, train.py:139) and upsample_iter=[3000,6000] is hardcoded β€” 6100/6000 is the shortest run that still exercises both resolution upsamples and the multi-scale -> single-scale transition.
  • --llffhold must match the scene's hold=N marker file (defocuscaps=5, blurcoffee=6).
  • Per-scene scripts/<scene>.sh use 10000-58000 iterations for full runs (min defocusseal 10000, max defocuscake 58000), not the README's flat
    1. defocuscaps.sh specifically uses 12000 with ms_steps=9000 plus two extra self-supervised regularizers (--use_depth_loss, --use_rgbtv_loss β€” total-variation on rendered depth/RGB, no extra data needed) not present in the README's example command; blurcoffee.sh is a clean 12000/ms_steps=6000 with no extras. Don't describe a smoke run as "the official config at fewer iterations" for defocuscaps β€” it's the README's simpler config, run short.
  • Unlike CoCoGaussian, metrics.py here runs standalone fine.
  • CoCoGaussian ships no scripts//configs/ at all, so cross-repo iteration comparisons should disclose which budget is author-documented (BAGS' per- scene scripts) vs. a bare code default (CoCoGaussian's iterations=30000 in arguments/__init__.py, uncorroborated by any README or config file).

Reproduce (smoke test, 6100 iters, README example config β€” not the per-scene official config):

./docker/run.sh python train.py -s data/defocuscaps -m deblurdata/defocuscaps \
    --init_dgt 0.0006 --iterations 6100 --ms_steps 6000 --min_opacity 0.1 \
    --eval -r 4 --port 6009 --kernel_size 0.1 --llffhold 5
./docker/run.sh python render.py  -m deblurdata/defocuscaps --data_device cpu --skip_train
./docker/run.sh python metrics.py -m deblurdata/defocuscaps -r 4 -t deblur-nerf

Results: defocuscaps PSNR 22.809/SSIM 0.625/LPIPS 0.307; blurcoffee_undist PSNR 27.044/SSIM 0.823/LPIPS 0.295. Full run: drop the iteration overrides and use the per-scene script's value.

Reproduce (full runs, official per-scene configs):

./docker/run.sh python train.py -s data/defocuscaps -m deblurdata/defocuscaps_official \
    --init_dgt 0.0006 --iterations 12000 --ms_steps 9000 --min_opacity 0.1 \
    --use_depth_loss --use_rgbtv_loss --eval -r 4 --port 6009 --kernel_size 0.1 --llffhold 5
./docker/run.sh python train.py -s data/blurcoffee_undist -m deblurdata/blurcoffee_undist_official \
    --init_dgt 0.0006 --iterations 12000 --ms_steps 6000 --min_opacity 0.1 \
    --eval -r 4 --port 6009 --kernel_size 0.1 --llffhold 6
# render.py + metrics.py exactly as above, pointed at the *_official model_path

No extra fixes needed beyond what's already here (Docker env, TLS certs, the blurcoffee undistortion). Results vs. paper (arXiv 2403.04926, Table 1): Caps 24.26/0.727/0.153 vs. paper's 24.16/0.725/0.159 (matches within noise); Coffee 30.63/0.903/0.117 vs. paper's 31.59/0.908/0.096 (~1dB low, plausibly because ours trained on the properly-undistorted scene while the paper likely used the distorted original).


Deblurring-3D-Gaussian-Splatting

Deblurring-3D-Gaussian-Splatting/ β€” one source patch (metrics.py). Python 3.12 venv via pip (no Docker needed).

Install.

git clone https://github.com/benhenryL/Deblurring-3D-Gaussian-Splatting.git --recursive
cd Deblurring-3D-Gaussian-Splatting
python3.12 -m venv deblurring_3dgs_env
./deblurring_3dgs_env/bin/pip install torch==2.4.1 torchvision==0.19.1 \
    --index-url https://download.pytorch.org/whl/cu121
./deblurring_3dgs_env/bin/pip install "numpy<2" plyfile lpips opencv-python-headless \
    scikit-image imageio configargparse tqdm

numpy<2 is required β€” the cu121 wheels otherwise pull numpy 2.x, which this codebase doesn't expect.

CUDA_HOME must be hand-assembled. Host CUDA (Ubuntu's nvidia-cuda-toolkit, 12.0) scatters files: nvcc at /usr/bin/nvcc, headers in /usr/include, libs in /usr/lib/x86_64-linux-gnu. /usr/lib/cuda looks like CUDA_HOME but its bin/ is empty, so the extension build fails to find nvcc. Build a symlink tree instead:

mkdir -p cuda_home/bin
ln -s /usr/bin/nvcc             cuda_home/bin/nvcc
ln -s /usr/include              cuda_home/include
ln -s /usr/lib/x86_64-linux-gnu cuda_home/lib64
ln -s /usr/lib/cuda/nvvm        cuda_home/nvvm

export CUDA_HOME=$PWD/cuda_home PATH=$CUDA_HOME/bin:$PATH
export CC=gcc-12 CXX=g++-12 TORCH_CUDA_ARCH_LIST=8.6 MAX_JOBS=8
./deblurring_3dgs_env/bin/pip install --no-build-isolation \
    ./submodules/simple-knn ./submodules/diff-gaussian-rasterization

gcc-12 is required (nvcc 12.0 rejects gcc-13). --no-build-isolation is required (else ModuleNotFoundError: No module named 'torch' β€” the submodules' setup.py imports torch directly).

Source patch β€” metrics.py (~line 86). Same root cause as CoCoGaussian's dead eval path but here it's a live crash, not dead code:

-                im1[i], im2[i], multichannel=True, full=True
+                im1[i], im2[i], channel_axis=-1, data_range=2.0, full=True

data_range=2.0 matches what the pinned scikit-image 0.19.3 inferred for the [-1,1]-rescaled floats this function receives, keeping numbers comparable to published results (omitting it lets modern skimage infer a different range per-image, silently shifting SSIM).

Dataset. real_camera_motion_blur/blurball (27 imgs, hold=7) β€” the scene in the repo's own README example.

Reproduce (smoke test, 3000/20000 iters):

export CUDA_HOME=$PWD/cuda_home
./deblurring_3dgs_env/bin/python train.py --expname smoke_ball \
  -s data/deblur_dataset/real_camera_motion_blur/blurball --config configs/real_camera_motion.txt \
  --iterations 3000 --test_iterations 200 1500 3000 --save_iterations 3000 --checkpoint_iterations 3000
./deblurring_3dgs_env/bin/python render.py -m output/smoke_ball \
  -s data/deblur_dataset/real_camera_motion_blur/blurball --iteration 3000

3000 iters was chosen to cross both densify_from_iter=500 and pts_iter=2500 (the paper's extra-point-allocation step). Result: test PSNR 26.418/SSIM 0.7479/LPIPS 0.3120, render.py reproduced the same PSNR to every digit from the saved checkpoint. Full run: drop the iteration overrides (default 20000).

Reproduce (full run, 20000 iters, official/default config):

export CUDA_HOME=$PWD/cuda_home
./deblurring_3dgs_env/bin/python train.py --expname full_ball \
  -s data/deblur_dataset/real_camera_motion_blur/blurball --config configs/real_camera_motion.txt

No iteration override β€” uses the config's default 20000. No extra fixes needed beyond the patch/env already documented above. Result: test PSNR 28.310 / SSIM 0.8267 / LPIPS 0.1492 (psnr.txt, [ITER 20000]), vs. paper's 28.27 / 0.8233 / 0.1413 for the Ball scene (arXiv 2401.00834) β€” very close match.

Other gotchas.

  • Don't relocate a dataset after training β€” cfg_args stores the absolute source_path.
  • environment.yml's pins (Python 3.7.13, torch 1.12.1, cudatoolkit 11.6) are unreachable via conda here and unnecessary β€” Python 3.12 + torch 2.4.1 works fine with the patch above.