Yoshino-master / FreeAnchor_TensorFlow

Tensorflow implementation for FreeAnchor: Learning to Match Anchors for Visual Object Detection, support for training your own dataset.

FreeAnchor_TensorFlow

1. Introduction

This is my implementation of FreeAnchor in pure TensorFlow. The original model is based on maskrcnn-benchmark. According to the paper "FreeAnchor: Learning to Match Anchors for Visual Object Detection".

2. Requirements

• python >= 3.5
• tensorflow-gpu >= 1.12.0
• pycocotools (if you want to train and eval on cocodataset)
• torchvision
• tqdm

3. Quick start on CocoDataset

• Download CocoDataset.
• Change the coco annotation. Change the directory in change_coo_data.py to your cocodataset path, and then run it.

python change_coo_data.py

• For training, just run the train.py file, the model will be saved in ./weight floder

python train.py

• For evaluating, I transfer the free_anchor_R-50-FPN_1x_8gpus.pth weights to tensorflow, you can download it and place it in ./weight floder. Then run the test.py file.

python test.py

4.Reference

@inproceedings{zhang2019freeanchor,
  title   =  {{FreeAnchor}: Learning to Match Anchors for Visual Object Detection},
  author  =  {Zhang, Xiaosong and Wan, Fang and Liu, Chang and Ji, Rongrong and Ye, Qixiang},
  booktitle =  {Neural Information Processing Systems},
  year    =  {2019}
}