- in pure tensorflow (include target generation, anchor generation, training, inference and post-process of result).
- use efficient tfrecord.
- ROI Align implement by tensorflow api with support of specific sample ratio.
- python3
- tensorflow == 1.12.0
- opencv-python
- scikit-image
python dataset_tools/convert_balloon_to_tfrecord.py --config_name=Config \
--num_shards=7 \
--num_workers=4 \
--split_name=train \
--dataset_dir=your_dataset_dir \
--tfrecord_dir=your_tfrecord_dir \
python train.py --checkpoint_dir=./_checkpoints --num_epochs=50 --gpus=0 \
--dataset_info=your_dataset_info_file \
--file_pattern=.../tfrecord/train*
python inference.py --checkpoint=your_checkpoint --gpus=0 --image_path=your_image_path
