Skip to content

models.md

Latest commit

 

History

History
36 lines (29 loc) · 1.15 KB

models.md

File metadata and controls

36 lines (29 loc) · 1.15 KB
Raw

Model Architecture Details

Available Models

The following models are available for use in this project:

1. UNet

  • Source: PyTorch implementation
  • Parameters: ~40M (depending on variant)
  • Backbone: ResNet-50 (pre-trained on ImageNet)
  • Encoder Head: Double convolution with skip connections

2. Deeplabv3

  • Source: PyTorch implementation
  • Parameters: ~69M
  • Backbone: ResNet-101 (pre-trained on ImageNet)
  • Encoder Head: ASPP (Atrous Spatial Pyramid Pooling)

3. PSPNet

  • Source: PyTorch implementation
  • Parameters: ~58M
  • Backbone: ResNet-50 (pre-trained on ImageNet)
  • Encoder Head: PSP Module (Pyramid Scene Parsing)

4. MANet

  • Source: Modified from original paper
  • Parameters: ~30M
  • Backbone: MobileNet-V2 (pre-trained on ImageNet)
  • Encoder Head: Light-weighted attention module

5. PAN

  • Source: PyTorch implementation
  • Parameters: ~140M
  • Backbone: ResNet-101 (pre-trained on ImageNet)
  • Encoder Head: Pyramid Attention Module

Each model is configured for semantic segmentation tasks and can be loaded with pre-trained weights or trained from scratch.