SeaFront - Synthetic dataset for visual container inspection

Applying DeepLearning techniques in a supervised manner generally requires a substantial amount of labeled data. However, such labeled data is not always readily available. To address this issue, one of the most common approaches is to synthetically generate  the necessary data for training the models. In the context of shipping container analysis,  an automatic synthetic image generation system  for containers in a port scenario has been created. This system can reproduce various visual aspects of interest, including the container itself from all sides, diverse realistic backgrounds,  potential damages the container might incur during shipment, IMDG stickers on the surface and text identification codes (BIC and ISO codes).  As a result, a database with automatically labeled images is obtained. This dataset has primarily two different objectives. The first one is to assist  researchers in training models capable of detecting the location  of the container in the image and the location and typology of the different elements that may be on the container's surface. The second one is to serve as ground-truth in evaluation tasks. We are making this dataset  publicly available, comprising almost 10000 images for training and validation, along with an additional  2480 images for testing. Our aim is to provide open and free data, which is often scarce in this field.

This resource contains two datasets:

1. Classification Task Dataset:
• Comprised of images categorized into two folders(door, no-door)
• Designed for a classification task, distinguishing between door and no-door images captured by C1 and C2 cameras.
2. Detection Task Dataset:
• Formatted in YOLOv5 format.
• Intended for detecting specific elements, including Text (BIC and ISO Codes), Containers, and IMDG stickers.