MSc dissertation project, Queen Mary University of London. Transfers human
motion capture (AMASS) onto unrigged animal meshes (DeformingThings4D) with
no skeleton, no rig, and no paired training data — predicts a per-face
Jacobian field, integrated by a differentiable Poisson solve.
Full write-up, all figures and tables, and the actual reported numbers:
see report_v6.docx / the results-overview notebook in the project
repository. This model card summarizes; it is not the source of truth.
⚠️ License — read before using
Research and non-commercial use only. These weights were trained in
part on AMASS, which is distributed under
an academic license restricting use to non-commercial scientific research.
Trained models inherit the most restrictive component license among their
training data (AMASS, DeformingThings4D). By downloading these
checkpoints you agree to:
Use them for research/non-commercial purposes only.
Not attempt to reconstruct the underlying AMASS or DeformingThings4D
data from these weights.
Cite the original AMASS, DeformingThings4D, and Neural Jacobian Fields
papers (see Citation below) alongside this work.
If your use case is commercial or you are unsure whether it's covered,
check the AMASS license terms
directly before proceeding — this summary is not a substitute for that
agreement.
Headline result — read this before trusting any single number
The project's central, best-supported finding is methodological, not a
model claim: the geometric rigidity metrics conventionally used to score
this kind of model (ARAP, smoothness) are confounded with how much the
mesh moves — a model that barely moves scores near-perfectly on them by
construction. Once displacement is controlled for, the ablation's headline
result (removing the perceptual loss "improves" rigidity, p=0.016) is
withdrawn — the two conditions occupy non-overlapping ranges of
motion, so the comparison is not identifiable. Full detail: report_v6.docx
Sections IV–V.
What is established: a working skeleton-free deformation substrate; a
genuine accuracy/rigidity trade-off when the ARAP term is removed (2.28x,
p<0.001); classical-ARAP metric parity (not improvement) against a
reproduced SfPT baseline; a diagnosed and partially fixed geometric
inversion-collapse failure mode.
What is not established: that the proposed perceptual loss (LVDP)
improves transfer quality — this is reported as a null result, not a
positive one.
Checkpoint inventory
All checkpoints are PyTorch state dicts: {"step", "model", "optimizer", "scheduler", "best_val", "cfg"} (encoder checkpoints have a different,
simpler schema — see below). Load with torch.load(path, map_location="cpu", weights_only=False); the code to actually run them
(MotionTransferModel, in src/models/pipeline.py) is in the project
repository, not included in this weights-only release.
The locked main result — warm-started from full_10k_s1337 with an inversion-barrier + motion-floor fix. This is the checkpoint the report's qualitative figures and demo use.
planC_lvdp_upright_s1337/best_val.pt
Up-axis-correction experiment — confirmed negative result (did not fix the underlying issue it targeted)
Up-axis fix applied to the canonical training path — reproduced the negative finding above
temporal_floor_1k_s1337/step_1000.pt
Frame-to-frame velocity floor experiment — spot-check looked positive, full 20-pair evaluation did not confirm it (frame-to-frame motion was significantly lower, not higher, under this term)
If you only want one checkpoint:fixAB_continue_s1337/step_800.pt is
the one the reported numbers and qualitative results in report_v6.docx
actually come from.
Data (not included, license-gated)
Running these checkpoints requires your own access to:
None are redistributed here or in the project repository, per their own
license terms.
Citation
If you use these weights, please cite the original datasets/methods this
project builds on, alongside this dissertation:
Mahmood, N. et al. (2019). AMASS: Archive of Motion Capture as Surface Shapes. ICCV.
Li, Y. et al. (2021). 4DComplete: Non-rigid Motion Estimation Beyond the Observable Surface [DeformingThings4D]. ICCV.
Aigerman, N. et al. (2022). Neural Jacobian Fields: Learning Intrinsic Mappings of Arbitrary Meshes. ACM TOG (SIGGRAPH).
Oquab, M. et al. (2023). DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193.
Reproducibility
Exact seeds (42, 1337), config, and evaluation protocol for every
checkpoint above are in report_v6.docx Table IV and the project
repository's START_HERE.md. python -m pytest -q (126 passed, 2
skipped) verifies the codebase these checkpoints were produced by.
