zhaoshiwen/blur-slam-bpn-data

Blur-SLAM BPN: pipeline-intermediate data + sample results (TUM fr1_desk, full-frame)

Self-produced data for the BPN deblur + 3D Gaussian/Triangle-Splatting pipeline in the companion code repo zhaoshiwen/blur-slam-bpn-code (see that repo's README for the full pipeline description and reproduction steps). This repo holds the pipeline-intermediate artifacts (EVSSM-deblurred frames, COLMAP reconstructions, training-ready scenes, depth maps) — everything needed to go straight to training without re-running EVSSM/COLMAP — plus one curated final result (TUM fr1_desk, full-frame, 613/613 frames COLMAP-registered, 50K iterations).

Raw source datasets (TUM RGB-D, ScanNet) are not included — download those from their official sources first; the dirs below are derived from them.

data/ — pipeline intermediates

evssm_deblurred_tum/<seq>/                 EVSSM-deblurred RGB frames, 1:1 with the
                                            raw TUM rgb/ sequence (e.g. fr1_desk: 613
                                            frames). Input to COLMAP.
evssm_deblurred_i2slam/<scene>/            same, for the i2slam-style scenes
                                            (ScanNet scene0024_01/scene0031_00/
                                            scene0736_00, TUM tum_fr1_desk[/_full]/
                                            tum_fr2_xyz/tum_fr3_office)

i2slam_colmap/<scene>/                     COLMAP reconstruction: images/ (symlinks
                                            to the deblurred frames) + sparse/0
                                            (cameras/images/points3D .bin+.txt from
                                            `feature_extractor` -> `sequential_matcher`
                                            -> `mapper` -> `model_converter`, all
                                            CPU). `tum_fr1_desk_full` = the 613/613
                                            full-frame reconstruction.

i2slam_da3/<scene>/                        Depth-Anything-V3 monocular depth maps
                                            (currently scene0024_01 only).

i2slam_trigsplat/<scene>_<variant>/        training-ready scenes: images/ (copied
                                            from i2slam_colmap), sparse/0/ (copied
                                            COLMAP model), depth/ (sparse-depth
                                            maps from gen_sparse_colmap_depth_generic.py),
                                            hold=8 (8 held-out test views). This is
                                            the `-s <DATA>` argument for train.py /
                                            train_bpn.py. `tum_fr1_desk_full_abl1`
                                            is the full-frame scene (613 images,
                                            613 depth maps, 1391.82 mean
                                            observations/image) used for the
                                            curated result below.

i2slam_trigsplat_gt/<scene>_<variant>/     GT-pose / GT-depth ablation variants of
                                            the above (TUM ground-truth trajectory
                                            instead of COLMAP pose, used for
                                            pose-source ablations).

<scene> values present: scene0024_01, scene0031_00, scene0736_00 (ScanNet), tum_fr1_desk, tum_fr1_desk_full, tum_fr2_xyz, tum_fr3_office (TUM RGB-D).

results/tum_fr1_desk_full/ — curated final result

Full-frame TUM fr1_desk (all 613 raw frames, 613/613 COLMAP-registered, -r 2, 50K iterations, --raw_blurry_stride 1, sharp-skip = top-15% NIMA-scored frames in tum_fr1desk_full_sharp_frames.json):

• tri_m2raw/ — triangle-splatting + BPN (--bpn_lr_kernel 4.5e-5 --bpn_lr_mask 4.5e-4). point_cloud/iteration_50000/point_cloud.ply (final trained representation), cfg_args, cameras.json, input.ply (initial COLMAP point cloud), renders/ (613 train-view renders at iter 50000).
• mip_m3eq/ — BAGS (Mip-Splatting) + BPN (--bpn_lr_kernel 4.5e-6 --bpn_lr_mask 4.5e-4, --max_shapes 2000000). Same layout, plus bpn_50000.pth (the trained per-frame BPN MLP weights).

To re-render or continue training, point -m at one of these directories and -s at data/i2slam_trigsplat/tum_fr1_desk_full_abl1 (use the code repo's render.py/train_bpn.py/train.py with the same flags as in cfg_args).

Expected local layout

The code repo's scripts expect this data under $BASE/data/... and write outputs to $BASE/outputs/..., e.g.:

$BASE/data/i2slam_trigsplat/tum_fr1_desk_full_abl1/   <- data/i2slam_trigsplat/tum_fr1_desk_full_abl1/ from this repo
$BASE/data/evssm_deblurred_tum/fr1_desk/              <- data/evssm_deblurred_tum/fr1_desk/ from this repo
$BASE/outputs/trigsplat_..._M2raw_full/tum_fr1_desk/  <- results/tum_fr1_desk_full/tri_m2raw/ from this repo (renamed)
$BASE/outputs/bags_..._M3eq_full/tum_fr1_desk/        <- results/tum_fr1_desk_full/mip_m3eq/ from this repo (renamed)

After cloning both repos, edit the BASE=/REPO_* path variables at the top of the scripts you run (see the code repo's README) to point at wherever you placed this data.