Turntable Marine Debris - Classification Task

This is convenience release containing data for the classification task of the Turntable dataset. Multiple datasets are available at various resolutions, from 32x32 to 96x96.

Images are cropped from the main marine debris turntable dataset and normalized at different pixel resolutions. The task is multi-class classification over 12 classes. Class names are provided in the 'class_names' dataset inside the HDF5 file. Each training set contains 1505 images, and the test set contain 645 images.

Water Tank Marine Debris - Classification Task

This is a sub-dataset containing data for the classification task. Multiple datasets are available at various resolutions, from 8x8 to 96x96.

Images are cropped from the main marine debris water tank dataset and normalized at different pixel resolutions. The task is multi-class classification over 11 classes, where the last class is background. Class names are provided in the 'class_names' dataset inside the HDF5 file. Each training set contains 1838 images, validation sets contain 394 images, and the test sets contain 395 images.

Water Tank Marine Debris with ARIS Explorer 3000

Marine Debris dataset captured at the Ocean Systems Lab Water Tank (Heriot-Watt University), using a ARIS Explorer 3000 Forward-Looking Sonar at 3.0 MHz frequency.

Classes included are: bottle, can, chain, drink carton, hook, propeller, shampoo bottle, standing bottle, tire, and valve. Bounding boxes are annotated in all images, with annotations stored in a JSON file.

Full sonar images projected in cartesian coordinates are available as PNG files. Image crops for each class are also included, which can be easily used with keras' ImageDataGenerator or other frameworks. These crops vary with size, so they need to be normalized before training a model with them.

Water Tank Marine Debris - Matching Task

This is a sub-dataset containing data for the matching task. Two datasets are available at 96x96 pixel resolution:

• Same (S): The same objects are used to build the training, validation, and testing sets. All object classes are used. Training set contains 33096 images, validation set contains 7092 images, and test set also 7092 images.
• Different (D): Objects of class 0-5 (can, bottle, drink carton, chain, propeller, and tire) are used to build the training set, while a different set of objects is used to build the test set (classes 6-9, hook, valve, shampoo bottle, and standing bottle). This is a harder task and requires object-independent learning to match FLS image patches. Training set contains 39840 images, while the test set contains 7440 images.

The task is given a sample pair of images, decide if they match or not, meaning that they contain the same object under a different point of view. This is modeled as binary classification. Each sample has shape (2, 96, 96), where the first dimension selects the sample from a matching pair. Additional labels (match_type) are available for each sample to indicate how it was obtained from the main marine debris water tank dataset:

• ID 0: Object to object positive, same class in the match pair.
• ID 1: Object to object negative, different class in the match pair.
• ID 2: Object to background negative

This information can be used for fine-grained evaluation and to find biases in the learned matcher.