Object Detection Model based on TensorFlow 1 Detection Model Zoo

Setup

Detectors are implemented with Tensorflow 1.15 and trained on NVIDIA GeForce RTX/GTX GPU devices with CUDA version 9 or 10.

Object detection model is trained with Tensorflow Object Detection API. Faster R-CNN with Resnet101 is selected from the Tensorflow Object Detection Model Zoo. The Faster R-CNN with Resnet101 backbone (faster_rcnn_resnet101_coco) is pre-trained on the COCO dataset, and is used to initialize our object detector model.

Configuring Spine ROI Detection Model

Our spinal regions of interest (ROI) detection models are trained with the following customizations:
- graph protos (*.pbtxt) for either axial (left, center or right) or sagittal
- config files (*.config) to generate the model graphs
- change the following
- num_classes
- batch_size
- fine_tune_checkpoint
- train_input_reader path
- eval_input_reader path

Train the Object Detection Model and Export Inference Graph

You can follow the guides and tutorials in the "References" below on TF1 object detection. Please make sure you are using a compatible commit of TF1 object detection repo.

We have provided two bash scripts to launch the training and exportation of trained graph:
- init training obj_det_train_py.sh
- export inference graph obj_det_export_inference.sh

Inspect Object Detection with Pickle

Object-Detection/object_detection.py is used to generate the detection pickle file.

References

• TF1 Object Detection API https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1.md
• TF1 Detection Model Zoo https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md
• TF1 Faster R-CNN with Resnet101 pre-trained on COCO dataset http://download.tensorflow.org/models/object_detection/faster_rcnn_resnet101_coco_2018_01_28.tar.gz
• Tutorial on TF1 object detection: https://pythonprogramming.net/introduction-use-tensorflow-object-detection-api-tutorial/