jiangguangai / fpn_faster_rcnn Goto Github PK

This tensorflow implementation code is based on Feature Pyramid Networks for Object Detection. **Some code learn from https://github.com/matterport/Mask_RCNN.

1. Streamlining structure and removing redundant parts
2. Making the code more clear
3. Replacing some API by new one
4. Adding the compute time graph by set the COMPUTE_TIME in config with True

FEATURES:

• Training >=2 image in a GPU
• Training uses Multi-gpu for data parallelism and no significant decrease in performance
• Add ResNet-v2 pretrained model
• The code has been tested and compared to this implementation, it has better performance
• Great debug tools and visualization to see result of intermediate process in the tensorboard
• Using tf.Dataset and tf.estimator which has better computational efficiency and more concise code

• tensorflow-1.8
• CUDA-9.0
• cuDNN-7.0
  The above environment has been tested.

for computational efficiency and coding convenient,we first convert the data into tfrecord format. We only support the transformation of COCO style data. If your data is the style of VOC, you must convert the VOC style to COCO style. One more thing you must notice is that you need to manually add the dictionary of category name and label in your data to /libs/label_dict.py. Then:

cd $FPN_Faster_RCNN/data/
python convert_data_to_tfrecord.py --DATA_dir=/cocodataset/ --annotation_dir=annotations/train.json --image_dir=images/train --save_dir=/tfrecord/ --dataset_name=coco --dataset_class=train

Generate:

save_dir

dataset_name

train.tfreocrd
test.tfrecord
val.tfrecord

this implementation has anthor feature is converted to change any params for users. I collect all params in config.py including: training params, network frame, and sample selection etc. So please change all params need to change for your training, such as: dataset dir, num_class, dataset_name, learning rate. To distinguish resnet50_v2 from resnet_v1, we named it as resnet_model. You can download the pre-trained model in there.

If you want train with multi-gpu, you just only change the GPU_GROUPS in config.py. Then:

cd tools/
python train.py

In the list, each file has a Debug signal to decide whether the corresponding summary are made in the tensorboard.

Of course, you can imitate our code to print out or visualate everything for debug.

Display images recorded in tensorboard:
ground true objects: proposals for training head: the first 50 proposal: the finial detection:

For convenience, our data analysis and the performance of the network are presented ipynb format.

• inspect_data.ipynb:Simple analysis and understanding of datasets and data.
• Predict.ipynb: predict the single image.
• evaluate_network.ipynb: compute mAP and the recall of rpn and error analysis.Error analysis including three types: classifier error, missing objects, false positive. And we also show some error objects to analysis.