# PAct Joint-Count Split Evaluation

Generated: `2026-05-16T06:32:50+00:00`
Portal: `http://106.14.105.96:28080/experiments/pact-joint-count-split-20260516/index.html`

Split rule: `few_joint` means GT movable joints <= 3; `many_joint` means GT movable joints >= 4.

| split | n | avg GT joints | avg pred joints | raw F1 | raw recall | count err |
|---|---:|---:|---:|---:|---:|---:|
| `few_joint <=3` | 74 | 1.84 | 1.45 | 0.4297 | 0.4054 | 0.45 |
| `many_joint >=4` | 26 | 10.00 | 3.58 | 0.1982 | 0.1729 | 6.42 |
| `very_many >=8` | 15 | 13.87 | 3.47 | 0.0863 | 0.0641 | 10.40 |

## Readout

Official PAct is much more reliable on few-joint/simple objects.  The main failure mode on many-joint objects is under-proposal: the predicted joint count grows slowly while GT joint count increases, so recall and F1 drop sharply.