Faster R-CNN R-101-FPN model was implemented with tensorflow eager execution.
- Cuda 9.0
- Python 3.5
- TensorFlow 1.11
- cv2
see train_model.ipynb
and inspect_model.ipynb
- evaluation utils
- training without for-loop when batch size > 1
- TTA (multi-scale testing and flip)
- online hard examples mining
- multi-scale training
- soft-nms
- balanced sampling and negative samples training (maybe)
This work builds on many excellent works, which include: