merve/SGinW

Dataset Card for Segmentation in the Wild

Dataset Description

Segmentation in the Wild (SegInW) is a computer vision challenge that aims to evaluate the transferability of pre-trained vision models. It proposes a new benchmark that assesses both the segmentation accuracy and transfer efficiency of models on a diverse set of downstream segmentation tasks. The challenge consists of 25 free, public segmentation datasets, crowd-sourced on roboflow.com, providing a wide range of visual data for model training and testing.

Composition

The SegInW challenge brings together 25 diverse segmentation datasets, offering a comprehensive evaluation of model performance across various scenarios. These datasets cover a broad range of visual content.

Data Instances

• Images: Visual data in the form of images, depending on the dataset.
• Annotations: Manual annotations specifying regions of interest or providing referring phrases for language-based segmentation.
• Segmentation Masks: Pixel-level annotations that define the boundaries of objects or regions in the visual data.
• Metadata: Additional information about the data, such as collection sources, dates, and any relevant pre-processing steps.

Data Splits Each folder has a train, train 10-shot and validation splits.

Dataset Creation The SegInW challenge is a community effort, with the 25 datasets crowd-sourced and contributed by different researchers and organizations. The diversity of sources ensures a wide range of visual data and evaluation scenarios. The datasets were labeled on roboflow.com as part of X-Decoder project.