search

dvlab-research / PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation (CVPR2021 Oral)
Apache License 2.0
391 stars 53 forks source link

All 400 model leads to 512 size dowload link #3

Closed lucasjinreal closed 3 years ago

lucasjinreal commented 3 years ago

Hi, it seems all 400 model size leads to a 512 size download link? Where to download a 400 model ?

yanwei-li commented 3 years ago

Yes. In this setting, we use 512 size for training and 400 size for testing (400 is too small when training for seg, which brings little performance drop).

lucasjinreal commented 3 years ago

@yanwei-li You mean I just need to set input resize to 400 and using 512 trained model?

yanwei-li commented 3 years ago

Yes. You could use the script PanopticFCN/configs/PanopticFCN-R50-400-3x.yaml and the pretrained model with 512-size training. Of course, it is also possible to retrain the model with 400 size, which will bring little performance drop. In our experiments, the gap between 400-size training and 512-size training (with 400-size testing) is about 0.2% PQ.

lucasjinreal commented 3 years ago

@yanwei-li thank u!