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
122 stars 21 forks source link

So,the STDCSeg is the best balance of speed and acc without distillation? #18

Closed ezio1320 closed 1 month ago

ezio1320 commented 3 months ago

Thank you for the nice code and comparsion of the real-time segmentation methods, Is the STDCseg is the best method? In other words, if i have a new segmentation job to do, Can i choose STDCSeg brainlessly？

（所以说STDC是目前最牛逼的实时语义分割方法哈？ 可以无脑用吧）

zh320 commented 1 month ago

It's hard to say which one is the best. As I said, the tested FPSs depend on many factors, such as different hardware, whether to use cudnn or not etc. In my experiments, without pre-training the backbone, STDC didn't reach reported results (<74%). Indeed, these performances were obtained by using deep supervision (i.e. auxiliary heads). In the same experiment setting, such as DDRNet, FarseeNet should achieve similar performance.

Best