Table 10 Sensitivity analysis of SST module parameters.
From: Learning spatio-temporal context for basketball action pose estimation with a multi-stream network
Training condition | Head AP (%) | Shou AP (%) | Wri AP (%) | Overall MAP (%) |
|---|---|---|---|---|
Baseline (no freezing) | 82.1 | 85.3 | 76.1 | 80.2 |
Freeze 4 layers | 82.3 | 85.5 | 76.3 | 80.4 |
Freeze 8 layers | 82.6 | 85.8 | 76.5 | 80.6 |
Freeze 20 layers | 81.9 | 85.1 | 75.9 | 80.0 |
High learning rate (0.01) | 81.5 | 84.7 | 75.6 | 79.8 |
Low learning rate (0.0001) | 82.8 | 85.9 | 76.7 | 81.1 |
Small batch size (16) | 82.0 | 85.4 | 76.0 | 80.1 |
Large batch size (128) | 82.5 | 85.7 | 76.4 | 80.5 |
