Model Zoo
Common Settings
We use distributed training with 8 GPUs by default.
Our ResNet style backbone are based on ResNetV1c variant, where the 7x7 conv in the input stem is replaced with three 3x3 convs.
There are two inference modes in this framework.
slide
mode: In this mode, multiple patches will be cropped from input image, passed into network individually. The overlapping area will be merged by average.whole
mode: In this mode, the whole imaged will be passed into network directly.
For input size of
8x+1
(e.g., 473, 769),align_corner=True
is adopted as a traditional practice. Otherwise, for input size of8x
(e.g., 512, 1024),align_corner=False
is adopted.
Supported Backbones
The supported backbones in SSSegmentation are summarized as following table,
Backbone | Model Zoo | Paper Link | Code Snippet |
---|---|---|---|
ConvNeXtV2 | Click | CVPR 2023 | Click |
MobileViTV2 | Click | ArXiv 2022 | Click |
ConvNeXt | Click | CVPR 2022 | Click |
MAE | Click | CVPR 2022 | Click |
MobileViT | Click | ICLR 2022 | Click |
BEiT | Click | ICLR 2022 | Click |
Twins | Click | NeurIPS 2021 | Click |
SwinTransformer | Click | ICCV 2021 | Click |
VisionTransformer | Click | IClR 2021 | Click |
BiSeNetV2 | Click | IJCV 2021 | Click |
ResNeSt | Click | ArXiv 2020 | Click |
CGNet | Click | TIP 2020 | Click |
HRNet | Click | CVPR 2019 | Click |
MobileNetV3 | Click | ICCV 2019 | Click |
FastSCNN | Click | ArXiv 2019 | Click |
BiSeNetV1 | Click | ECCV 2018 | Click |
MobileNetV2 | Click | CVPR 2018 | Click |
ERFNet | Click | T-ITS 2017 | Click |
ResNet | Click | CVPR 2016 | Click |
UNet | Click | MICCAI 2015 | Click |
Supported Segmentors
The supported segmentors in SSSegmentation are summarized as following table,