Pose Embed
How do you recognize an action you have never learned—when the only clue is a damaged skeleton?
Follow the experimentThe action is unfamiliar.
The representation doesn’t have to be.
MotionBERT turns a sequence of body joints into a motion signature. We freeze that encoder and train only a linear head: a new lens for comparing the signatures.
Two people · 17 joints · x, y, confidence
The same representation for every recipe
Normalize, compare by cosine, retrieve
At test time, one clean exemplar represents each of 20 unseen actions. A query retrieves the most similar exemplar. That is the one-shot task.
Does the neighborhood around a motion help us recognize it under damage?
Standard contrastive training works with pairs. The full contextual recipe also compares their neighborhoods. We test whether the complete recipe loses less accuracy when queries are corrupted.
The comparison has a boundary. Margins and regularization differ too, so an advantage would support the full recipe—not isolate the neighborhood term. SupCon is a descriptive comparator.
Keep the gallery clean.
Change only the query.
Three separate stress tests, each at three nested severities. Every recipe sees the same damage. Try the query controls in the illustration.
Perturb visible joint locations.
Mask the same selected joints across the sequence.
Mask a consecutive block without deleting time.
These are controlled tensor stress tests. They do not establish robustness to real camera occlusion or a different pose estimator.
Fourteen weeks. Each step produces the evidence the next one needs. Explore any week, or keep scrolling.
Week 1 / Sep 15-20
ClosedWhat would count as a trustworthy answer?
Specify the retrieval question, verify licensed poses and the pretrained checkpoint, and measure the frozen encoder on an allocated GPU.
A hash-bound starting protocol, an input inventory, and measured compute evidence.
Why it matters. Agreeing on the rules before seeing test outcomes keeps a disappointing answer just as valid as an encouraging one.
These are recorded tracker states. The calendar alone never marks work complete.
Week 2 / Sep 21-27
ClosedCan we recognize an action we never trained on?
Assign fixed identities to every sample. Separate 80 training actions, 20 development actions, and 20 unseen test actions; keep one clean test exemplar per action.
Seven reproducible manifests and a leakage audit, including exact lists of excluded camera views.
Why it matters. The primary queries exclude every synchronized view of the gallery performance, so another camera cannot give away the answer.
These are recorded tracker states. The calendar alone never marks work complete.
Week 3 / Sep 28-Oct 4
ClosedCan every result be traced back to its exact inputs?
Version the frozen MotionBERT path for v3, verify retrieval and cache provenance, audit historical test access, and freeze the design before fresh extraction.
A verified auxiliary-only preparation path with a fixed design, measured torso fallback, and registered artifact roots.
Why it matters. Features, code, and protocol must agree. A new version name cannot erase prior test access or turn historical evidence into v3 results.
These are recorded tracker states. The calendar alone never marks work complete.
Week 4 / Oct 5-11
ClosedAre we training the mathematics we intended?
Verify loss values and gradients, paired initialization, and physical batches of eight actions × four examples. Generate fresh features and run one complete 20-epoch pilot per recipe.
Three pilot records with checkpoints at epochs 5, 10, 15, and 20, plus numerical health, timing, memory, and storage evidence.
Why it matters. Pilots test implementation and capacity. Their scores cannot pick learning rates, change the recipes, or select a shorter training duration.
These are recorded tracker states. The calendar alone never marks work complete.
Week 5 / Oct 12-18
ActiveCan the entire study finish without changing its rules?
Verify all nine development corruptions and the analysis fixtures. Measure the complete extraction-to-retrieval path, then check GPU access, available time, and storage.
An October 18 go/no-go decision with at least 25% work slack and a protected 200 GiB storage reserve.
Why it matters. The full campaign must fit before selection begins. A failed gate calls for a documented result-blind amendment, not a smaller budget for one recipe.
These are recorded tracker states. The calendar alone never marks work complete.
Week 6 / Oct 19-25
PlannedWhich learning rate works on clean, held-out actions?
Begin 27 fresh development runs: three recipes × three learning rates × seeds 7, 17, and 29. Pair the initial heads and exact physical batches within each seed.
Immutable attempts and clean development scores measured at the fixed epoch-20 endpoint.
Why it matters. Equal attempted budgets make the comparison interpretable. Corrupted development scores and pilot weights stay outside selection.
These are recorded tracker states. The calendar alone never marks work complete.
Week 7 / Oct 26-Nov 1
PlannedCan another reader reproduce our selection?
Finish all 27 attempts. Rank each recipe’s valid three-seed candidates by mean clean top-1, then mean reciprocal rank, then the lower learning rate.
Three timestamped learning-rate choices and paired plans for final training on all 100 auxiliary actions.
Why it matters. A candidate needs all three valid seeds. Keeping failures visible prevents a lucky partial result from becoming the selected configuration.
These are recorded tracker states. The calendar alone never marks work complete.
Week 8 / Nov 2-8
PlannedWhat geometry does each recipe learn?
Start nine fresh linear heads on all 95,001 auxiliary examples. Keep MotionBERT frozen; train each recipe with the three paired seeds for exactly 20 epochs.
Final training attempts with recorded initialization, batch plans, health diagnostics, and resource costs.
Why it matters. Each head maps the same 8,704 features to 2,048 normalized values. Shared inputs let us compare the full training recipes under paired conditions.
These are recorded tracker states. The calendar alone never marks work complete.
Week 9 / Nov 9-15
PlannedAre all nine models ready for the same test?
Complete the nine final runs and audit their checkpoint hashes, exact update counts, and provenance. Declare both query definitions and every evaluation condition.
Nine verified epoch-20 heads and a fixed matrix of 180 expected result cells.
Why it matters. Pilot or development weights cannot substitute for final training. Every recipe and seed must be ready before unseen test evaluation begins.
These are recorded tracker states. The calendar alone never marks work complete.
Week 10 / Nov 16-22
PlannedWill evaluation measure exactly what we promised?
Independently check rankings, ties, query subsets, all-head projection, and complete result accounting using fixtures and development data. Rehearse figures and reports.
A reviewed evaluator, verified opening controls, and report templates explicitly marked as non-final.
Why it matters. We can fix a scoring defect before opening. After test outcomes are visible, changing the confirmatory pipeline would change the scientific claim.
These are recorded tracker states. The calendar alone never marks work complete.
Week 11 / Nov 23-29
PlannedIs everything fixed before the unseen actions appear?
Resolve audit findings and bind all nine checkpoints, exact sample lists, code, and evaluation rules into the final lock. Repeat seal and capacity checks by November 29.
A complete authorization package, with Thanksgiving time reserved for recovery and any remaining blockers made explicit.
Why it matters. Training must be finished, historical access resolved, and resources available. An unresolved prerequisite keeps the test sealed.
These are recorded tracker states. The calendar alone never marks work complete.
Week 12 / Nov 30-Dec 6
PlannedHow much recognition survives damaged pose queries?
Record the authorized opening before loading novel poses. Keep the gallery clean; evaluate clean queries and nine corruption conditions through every final head and both query definitions.
180 provenance-verified cells with top-1, reciprocal rank, top-5 retrieval, and reproducible per-query ranks.
Why it matters. All heads reuse the same encoded inputs and realized damage. A crash permits only the identical authorized retry; it never reseals the test.
These are recorded tracker states. The calendar alone never marks work complete.
Week 13 / Dec 7-13
PlannedDoes the full contextual recipe lose less accuracy?
Compare clean-to-corrupted top-1 drops across the three paired seeds and nine conditions. Compute the prespecified conditional interval, then show absolute scores and every seed and cell.
Reproducible curves, action sensitivity checks, resource tables, and a report and poster draft.
Why it matters. Smaller degradation alone can hide poor accuracy. A contextual advantage requires the interval’s upper bound below zero; null and contrary outcomes remain successful findings.
These are recorded tracker states. The calendar alone never marks work complete.
Week 14 / Dec 14-18
PlannedCan someone follow the evidence all the way back?
Run an independent reproduction smoke test, finish the report and poster, and archive permitted code, configurations, manifests, and compact summaries by December 18.
A reproducible final study with explicit limits—or an honest incomplete-study record if the required campaign could not pass its gates.
Why it matters. The deliverable is a defensible answer, not a promised win. Licensed poses, checkpoints, full run artifacts, and private records stay outside the public archive.
These are recorded tracker states. The calendar alone never marks work complete.
We expect contextual neighborhoods may help preserve retrieval under damage. The experiment is designed to find out—not to make that expectation come true.
The prespecified conditional interval lies below zero. We can support a smaller accuracy drop for the full contextual recipe.
The evidence does not resolve an advantage. We report the uncertainty, absolute accuracy, and every declared condition.
The result runs against the hypothesis. We document where it fails and what the comparison can—and cannot—explain.
Δ = contextual accuracy drop − contrastive accuracy drop
Negative means less degradation. A smaller drop alone does not mean better corrupted accuracy.Three pilots. Twenty-seven selection runs. Nine final heads.
180 result cells. One defensible conclusion.