Creating a set of .png images and their associated masks (one png mask per worm) from an HDF5 file and XML hand-drawn annotations.
For this task, use the file annotation_data_generation.py
Supplementary training data, with individual worms added in to images to mimic the conditions of the failure cases of the Tierpsy tracker, can be generated within the synthetic_data_generator.ipynb python notebook.
The original images as recorded by the camera are 2048x2048 pixels. Training on these images exceeds the memory capacity of a typical Titan X GPU (11GB). As such, for this investigation the images were cropped into 16 smaller tiles of 512x512px each.
This is automatically performed by the annotation_data_generation.py file, but for training on existing data this can be performed with the data_cropper.py file.
The TensorFlow models used for training take data in the tf record file format. The tf_data_creation_loop.py file will run through a set of datasets, creating a tf record "shard" of training and validation data from each set.
Models can be trained on a list of shards, specified from within the config file.
Full instructions for training can be found at https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md
It is recommended to perform fine-tune training from a model pre-trained on COCO as full training can take up to 2 weeks. Pre-trained checkpoints can be found at https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
The object detection API includes an evaluation script (eval.py, replicated in this repository) which can be used to evaluated the trained model against the COCO evaluation metrics of mAP and mAR.
Once the model has been trained, the graph can be exported for inference.
The exported inference graph can be used in either the inference.py script, or inference_nb.ipynb Python Notebook.
The results should look something like this:
