search

zh320 / realtime-semantic-segmentation-pytorch

PyTorch implementation of over 30 realtime semantic segmentations models, e.g. BiSeNetv1, BiSeNetv2, CGNet, ContextNet, DABNet, DDRNet, EDANet, ENet, ERFNet, ESPNet, ESPNetv2, FastSCNN, ICNet, LEDNet, LinkNet, PP-LiteSeg, SegNet, ShelfNet, STDC, SwiftNet, and support knowledge distillation, distributed training etc.
Apache License 2.0
107 stars 19 forks source link

LiteSeg model only has one concate but the model in paper has two #14

Closed zhy168 closed 4 months ago

zhy168 commented 4 months ago

And your mIoU is 76% that is much higher than 66%-70% in paper.

best wishes

zh320 commented 4 months ago

Hi zhy,

There are two concatenate within the codes, one is in this line

https://github.com/zh320/realtime-semantic-segmentation-pytorch/blob/0e5fe540f13c91d36ec30e745dc151ad356d3fd4/models/liteseg.py#L39

The other one is inside the DASPPModule

https://github.com/zh320/realtime-semantic-segmentation-pytorch/blob/0e5fe540f13c91d36ec30e745dc151ad356d3fd4/models/liteseg.py#L72

Regarding the improvement of performance, it's probably due to tricks like train more epochs, more augmentations and using ema.

Best,

zhy168 commented 4 months ago

Thanks a lot