A PyTorch implementation of PSPNet18 with pixelwise knowledge distillation.
This repo is based on and modified from semseg.
Crop size 224 * 224
With 713 * 713 pretrained PSPNet50 as teacher_net:
Main_loss = Kldiv * (1 - alpha) + cross_entropy * alpha
| Config | mIoU/mAcc/aAcc |
|---|---|
| T1.8 A0.2 | 0.7130/0.7904/0.9505. |
| T2 A0.0 | 0.7119/0.7876/0.9507. |
| T2 A0.1 | 0.7116/0.7876/0.9501. |
| T2 A0.2 | 0.7129/0.7895/0.9503. |
| T2 A0.5 | 0.7095/0.7863/0.9504. |
| T3 A0.1 | 0.7023/0.7820/0.9500. |
| T4 A0.2 | 0.7072/0.7847/0.9500. |
| T6 A0.8 | 0.7045/0.7800/0.9494. |
| T10 A0.2 | 0.6981/0.7747/0.9488. |
| Baseline481 | 0.6927/0.7639/0.9483. |