runs at lowest Foxhole graphics settings with 1024 x 768 resolution
This is way more complicated than it needs to be.
https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/install.html
- to update labels.csv used by training records run py -m image_recognition.xml_to_csv
- update unique labels in CLASS_MAP in image_recognition/generate_tfrecord.py
- to create training records py -m image_recognition.generate_tfrecord --csv_input=image_recognition/annotations/train_labels.csv --image_dir=images/screens/1024-768 --output_path=image_recognition/tfrecords/train.record --class_descriptor_output_path=image_recognition/tfrecords/class-descriptor.pbtxt py -m image_recognition.generate_tfrecord --csv_input=image_recognition/annotations/test_labels.csv --image_dir=images/screens/1024-768 --output_path=image_recognition/tfrecords/test.record --class_descriptor_output_path=image_recognition/tfrecords/class-descriptor.pbtxt
- train model py -m object_detection.model_main --logstdserr --train_dir=image_recognition/tfrecords/ --pipeline_config_path=image_recognition/trained_models/faster_rcnn_inception_v2_foxhole/model.config
- view training progress tensorboard --logdir=image_recognition/tfrecords/
- when satisfied with model generate trained inference graph python -m object_detection.export_inference_graph --input_type image_tensor --pipeline_config_path image_recognition/trained_models/faster_rcnn_inception_v2_coco/pipeline.config --trained_checkpoint_prefix model/model.ckpt-200 000 --output_directory image_recognition/inference_graph
- run live image detection. your primary monitor must be set to foxhole and on 1024x768 res. py -m image_recognition.game_image_detection