Feature Pyramid Network is ubiquitious for detection, semantic segmentation, or even panoptic segmentation, still found in state-of-art models. Most implementations around are done in PyTorch, here I made my own spin on FPN in keras for usage in TF-based projects.
Main objective of creating this project is to compare FPN performance with different backbones for object detection.
Main code can be found in my jupyter notebook trained and ran.
Please check new_model.pyfor how the model is defined.
All the output layers from FPN is flattened and concatenated so there's only 1 output. This is done so that it can be connected to a final Dense layer(s) for cifar10 predictions. You may wish to remove this head for your purposes.
Bottom-up pathway in my implementation is compatible with:
- ResNet50
- ResNet101
- ResNet50V2
- ResNet101V2
Feel free to add compatibility with other backbones! Possible backbones to add:
- Wide ResNet
- ResNeSt
- ResNeXt
- Models with Res2Net
Original FPN paper: https://arxiv.org/abs/1612.03144 Panoptic FPN paper: https://arxiv.org/abs/1901.02446