Cloth-splatters/dexgarmentlab-folding-lifting-meshes-clean

DexGarmentLab folding/lifting meshes — point-cloud cleaned

A segmented copy of Cloth-splatters/dexgarmentlab-folding-lifting-meshes with non-cloth points removed from the observations. Ground-truth mesh positions, topology and gripper poses are bit-identical to the source; only the pointclouds datasets differ.

⚠️ Read this before using it

The segmentation uses the ground-truth cloth mesh to decide which points to keep. That makes this dataset:

• ✅ correct for training — the supervision signal was always available at training time, so using it to clean the inputs leaks nothing;
• ❌ wrong for evaluation — a state estimator benchmarked on these clouds is handed a segmentation it would not have at inference, which flatters it.

For evaluation, use the original dataset together with an outlier-robust observation likelihood. The UniClothDiff particle filter exposes obs_trim_fraction for exactly this: it discards the worst-fitting fraction of observed points, needs no ground truth, and therefore also covers real sensor clutter.

Why this exists

The merged point clouds in the source dataset contain static scene geometry — a rig roughly 0.8 m above the table, 0.30–0.45 m away from the cloth — that no cloth mesh can explain.

Measured against the ground-truth mesh at a 0.05 m threshold, across all 239 cloths:

contamination (mean fraction of points > 5 cm from the cloth)	cloths	share
clean (< 1%)	123	51.5%
1–10%	17	7.1%
10–25%	43	18.0%
25–50%	21	8.8%
50%+	35	14.6%

Worst cloth: TCNC_Top063_obj at 74.7% mean (80.4% peak). Overall mean 15.9%.

This has two consequences. Any likelihood of the form "every observed point should be near some mesh vertex" acquires a large constant offset that is identical for every hypothesis, which destroys discrimination between them. And because the released GPS state-estimation checkpoints were trained on these clouds, the contamination is baked in upstream of anything downstream.

Concretely, in a test that reconstructs one observation under all 239 candidate meshes and ranks the true one:

observation	true-mesh rank (of 239)
clean cloths (5)	1, 1, 1, 1, 1
contaminated cloths (3)	10, 51, 24

The likelihood is near-perfect on clean observations and near-chance on contaminated ones.

Method

For every frame and camera, a point is kept when it lies within 0.05 m of the nearest ground-truth cloth vertex at that frame. Frames retaining fewer than 256 points instead keep their 256 closest, so a badly segmented frame degrades rather than producing an empty cloud (this did not trigger on any frame here).

Simple background subtraction was considered and rejected: only 10–30% of non-cloth points recur at identical coordinates between consecutive frames, because the scene surfaces are re-sampled every frame.

Contents

Identical structure to the source: training/<cloth>/<trajectory_N>/step_M/ with positions, gripper_pos, and pointclouds/cam_0.

• 239 cloths, 894 trajectories, 28 673 frames — all cleaned, none subsampled
• points retained: mean 90.2%, 10th percentile 67.3%, minimum 19.1%
• clouds are now variable-length per frame (the source is a fixed 2048); resample_points_with_mask pads and masks them for the encoder

Reproducing

python scripts/clean_pointclouds.py \
    --h5 dexgarmentlab_folding_lifting_meshes.h5 \
    --output dexgarmentlab_folding_lifting_meshes_clean.h5 \
    --threshold 0.05

Audit the source (or any other trajectory dataset) with:

python scripts/audit_pointcloud_contamination.py --h5 <dataset>.h5

Both scripts are in UniClothDiff.