Faster R-CNN is a really classical two-stage framework for object detection. Common tensorflow versions are often in complicated class implementation. For beginner or fresher, reading those codes is very hard. So I build this repo reproducing the Faster R-CNN by tensorflow without any complicated class, just functions and sufficient comments. It is easy to read and understand. The project can guide you to build your own Faster R-CNN framework. Hope you like this. You also can give any advise for this repo, welcome to technical exchange (as well as my English mistakes). :)
- No class implementations
- Sufficient comments
- Easy to read and understand
- Tensorboard shows rpn and rcnn results
- Training in a toy dataset, no download large datasets
- Training in large dataset(late support)
Tensorflow >= 1.11 opencv >= 3.4.0
- Batch size == 1
python train.py
- Batch size > 1
python train_batch.py
- COCO is not available.
python test.py
tensorboard --logdir=./logs
- Add references.
- Add MobileNet.
- Training with other datasets as COCO or VOC.