This repository contains the dataset and some of the configuration files to train an object detector of R2-D2 and BB-8 from Star Wars, using Tensorflow's object detection API.

The final model has been fine-tuned using one of the models pre-trained on the COCO dataset as a starting point (transfer learning).

You can read more here.

Here there's a video to show the performance of the model.

The outcome seems acceptable considering the small size and low variability of the dataset used.

In order to try the dataset, you first need to follow the installation instructions on the Tensorflow page. Then, you can train your own model following analogous commands as the ones used in the pet detector.

• annotations: xml files with the boundig boxes of each image, obtained using labelImg, and a text file assigning each image to one of the two available classes (R2-D2 and BB-8). Analogous format as the one of the Oxford-IIIT Pets Dataset.
• graph_images: plots of the Total Loss of the trained model.
• images: 203 jpg train images, 102 of R2-D2 and 101 of BB-8.
• object_detection/data: configuration file of the net trained, label map and tensorflow records (created with create_sw_tf_record.py).
• results: some images/gifs showing the performance of the trained model.
• test_images: some images not used to train the model.

• check_duplicate_images.py: script that allows you to check if there are duplicate images in one or two directories. Useful when collecting the train and test images.
• create_sw_tf_record.py: modified version of create_pet_tf_record.py, where the paths have been changed along with the regular expression in line 60 to adapt to my images names.
• export_inference_graph.py: an exact copy of export_inference_graph.py in Tensorflow's API.
• object_detection_sw.ipynb: ipython notebook to test the trained model on images and videos. Modified from object_detection_tutorial.ipynb and some lines of code to test videos from priya-dwivedi github (I don't know if that's the original source).
• rename_images.py: script to rename all images files in one directory to have the same root name.