# PAct Complex Axis Head Training

Generated: `2026-05-16T09:04:26+00:00`
Portal: `http://106.14.105.96:28080/experiments/pact-complex-axis-head-train-20260516/index.html`

This is a lightweight learned proxy for the proposed Stage-2 instance-conditioned articulation head.  It trains/evaluates an axis candidate ranker on complex GAPartNet samples using leave-one-object-out evaluation.

| split | parts | majority accept | learned accept | global xyz oracle | candidate oracle | majority err | learned err | oracle err |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| `few_joint <=3` | 0 | 0.0000 | 0.0000 | 0.0000 | 0.0000 | 0.00 | 0.00 | 0.00 |
| `many_joint >=4` | 259 | 0.7297 | 0.7915 | 0.8996 | 0.9498 | 22.69 | 16.43 | 5.97 |
| `very_many >=8` | 208 | 0.7837 | 0.8462 | 0.8942 | 0.9423 | 18.21 | 11.99 | 6.97 |

## Readout

Training on complex data is feasible and useful as a lightweight head: the learned ranker should outperform global-axis priors if semantic/root-frame/local-geometry features contain the missing signal.  The oracle column estimates the ceiling of the current candidate set.