This repository contains the source code and the dataset related to the paper: F. Zocco, T. -C. Lin, C. -I. Huang, H. -C. Wang, M. O. Khyam and M. Van, "Towards More Efficient EfficientDets and Real-Time Marine Debris Detection," in IEEE Robotics and Automation Letters, vol. 8, no. 4, pp. 2134-2141, April 2023, doi: 10.1109/LRA.2023.3245405.
An overview of the main folders is below. For details, see the .txt files inside each folder.
The folder "_modified-EfficientDet" is a modification of the original TensorFlow implementation of EfficientDets, version of 23 October 2021: https://github.com/google/automl/tree/master/efficientdet. The description of the architectural modifications is in the paper mentioned above. Here there is its implementation. We found that the architecture with less BiFPN layers and deeper class/box nets is more efficient.
Files where I made changes or files that I have added (I commented with my name the changes in each file):
(1) re-wrote "tutorial.ipynb" (resulting in "Tutorial-training_PASCAL-federico_zocco.ipynb", "Tutorial-training_TrashICRA19-federico_zocco.ipynb", "Tutorial-inference_images-federico_zocco.ipynb" and "Tutorial-model_eval-federico_zocco.ipynb")
(2) slightly modified "hparams_config.py"
(3) slightly modified "model_inspect.py"
(4) created "dataset/prepare_xml-federico_zocco.py"
(5) added "dataset/generate_tfrecord.py"
(6) created "dataset/generate_tfrecords_Colab-federico_zocco.ipynb"
(7) slightly modified "dataloader.py"
(8) slightly modified "tf2/label_util.py" to add the Trash-ICRA19 and WPBB labels map
(9) slightly modified "inference.py" to make either the WPBB or the Trash-ICRA19 labels map the default one
(10) added the folder "inference_images".
See each file for details.
The folder "WPBB_dataset" contains the in-water plastic bags and bottles (WPBB) dataset, which currently has 900 fully annotated images for object detection. Annotations are in PASCAL VOC 2012 format.
The folder "Low-light_detection_Results" shows how the detection results in Table IX of the related paper were achieved. We found that, for real-time robotic applications, the simple low-light image enhancement method based on adding a constant value to all the pixels remains a competitive approach.
