aLongZJU/DirtyMoCap-Evaluation

DirtyMoCap Evaluation

This benchmark contains 414 paired body/hand motion sequences under six marker/noise conditions: 2484 sequence inputs in total. The six conditions reuse the same motions; they are not 2484 distinct actions. All frames are retained. Each condition is a separate evaluation, not a training split.

Files

README.md
release.json
SHA256SUMS
reference_metrics.json
evaluation/
  localmocap_missing/
    manifest.json
    <sample_id>.npz
  openmocap_missing/
  soma_missing/
  localmocap_occ_outlier/
  openmocap_occ_outlier/
  soma_occ_outlier/

Condition	Sequences	Frames	NPZ size (GiB)
localmocap_missing	414	581,878	1.541
openmocap_missing	414	581,878	1.592
soma_missing	414	581,878	1.574
localmocap_occ_outlier	414	581,878	1.508
openmocap_occ_outlier	414	581,878	1.559
soma_occ_outlier	414	581,878	1.541

NPZ files and per-condition manifests are byte-for-byte copies of the data used in the completed evaluation. Noise, sequence lengths, sample order, saved permutations and ground truth have not been regenerated. release.json lists file sizes and SHA256 hashes; SHA256SUMS also covers the release index. Neither file contains its own hash.

From this dataset directory, verify the files with:

sha256sum --check --quiet SHA256SUMS

Model weights, SMPL-H/MANO model files, predicted meshes, local logs and pickled training batches are not included.

NPZ format

All coordinates are float32 in meters, in the sequence world coordinate frame. Each frame has 173 padded marker slots. Missing observations must be interpreted using valid_mask; a zero coordinate alone is not a validity test.

Array	Shape	Meaning
markers	[T,N,3]	Saved noisy observations in their original saved point order
valid_mask	[T,N]	Boolean validity for the marker slots
shuffle_perms	[T,N]	Saved permutation of all marker slots, separately for each frame
gt_points	[T,113,3]	Saved GT: 52 joints followed by 61 surface anchors
gt_poses	[T,156]	Axis-angle poses: global orientation 3, body 63, left hand 45, right hand 45
gt_transl	[T,3]	GT translation
gt_betas	[10] or [T,10]	GT SMPL-H shape coefficients
anchor_ids	[61]	Surface vertex indices, matching DirtyMoCap's anchor configuration
frame_ids	[T]	Source frame numbers
fps	scalar, optional	Playback rate, only where source metadata has been verified
format, unit	string scalars	dirtymocap.evaluation.v1 and m
metadata	JSON string scalar	Sample identity, condition, source motion and marker-order information

Load without pickle:

import numpy as np

with np.load("evaluation/<condition>/<sample_id>.npz", allow_pickle=False) as data:
    markers = data["markers"]
    valid_mask = data["valid_mask"]
    permutation = data["shuffle_perms"]
    markers = np.take_along_axis(markers, permutation[..., None], axis=1)
    valid_mask = np.take_along_axis(valid_mask, permutation, axis=1)

Apply the saved permutation exactly once to both observations and validity. DirtyMoCap's evaluator does this automatically; do not pre-shuffle its input files. Valid points remain unordered and have no provided anatomical correspondence. GT arrays are scoring targets and must not be passed to the inference network.

Evaluation

Using the DirtyMoCap repository and its prepared environment, body model and paired v1 weights, set --data-root to this package's evaluation directory. For example, if this package is stored at assets/evaluation-release, run from the code repository root:

python tools/evaluate_suite.py \
  --data-root assets/evaluation-release/evaluation \
  --output-dir workspace/evaluation/small \
  --indices 0 263 295 311 --gpus 0

This evaluates four complete motions per condition. Remove --indices for all 2484 inputs. Add --mode refine for Refine-only evaluation. See the code repository's Install.md and Evaluation.md for setup, GPU memory settings, metrics and resume behavior.

The reference run uses first-frame Init, seed 0 separately for each condition, and continuous RNG consumption in manifest order. Native sparse/voxel operations can introduce numerical variation. A small subset is useful for checking setup; it does not reproduce the full run's RNG history for later samples.

Condition	Refine 52 joints (mm)	Solver 52 joints (mm)	Solver MPVPE (mm)
localmocap_missing	7.206603	7.896252	9.203616
openmocap_missing	8.583396	8.855341	9.133885
soma_missing	6.596706	6.950956	7.572744
localmocap_occ_outlier	9.122993	9.825386	10.424221
openmocap_occ_outlier	8.990407	9.282382	9.502515
soma_occ_outlier	7.556984	7.906242	8.151804

Errors use world coordinates without root alignment or rigid registration. First average over points and frames within each sequence, then average equally over sequences. MPVPE reconstructs both predicted and GT meshes with full SMPL-H pose correctives. Saved GT anchor coordinates retain their original generation geometry. Full-precision metrics, manifest identities and checkpoint source fingerprints are recorded in reference_metrics.json.