FAR – Free Form Object Centric Captioning

What kind of information is provided to the annotators about the images or videos?

• Only the position of the objects.

• Only the segmentations of the objects.

• Damage/no-damage label, object segmentation masks, and corresponding paragraphs.

• No additional information.

What are the steps that need to be followed to create an annotation of an item from scratch?

• Draw a tight bounding box over the item and add a description of the item, focusing on attributes like color, shape and functionality.

• There is no need to create an annotation of the item region, only validate and improve the existing ones.

• Only draw a tight bounding box over the item.

• Only add a description of the item, focusing on attributes like color, shape and functionality.

  
  

To annotate the damaged region of an item, the following steps should be followed:

1. 1- Draw a tight bounding box over the damaged/abnormal region of the item.

2. 2- Add a description of the abnormality, focusing on but not limited to the type of damage, package deconstruction, deformation, impact on functionality.

• Verdadero
• FALSO

What is the "Text Captioning" provided and what is its purpose?

• "It is a previously generated model that describes the item and scene, including abnormalities (if any).

• It is a descriptive caption that should be added to the annotation of the item region.

  
  

• It is a text that should be ignored, as the annotations must be created from scratch.

  
  

• It should be used as a reference and copied/pasted into the text description of the item annotation, if possible."

"What types of problems or conditions are considered "damage" to the items according to the information provided?"

• Items that are opened or unpacked

  
  

• Items that are crushed or deformed

  
  

• Items with tears, holes, or broken seals

  
  

• All of the above

Do you need to produce free-form captions for each item, even if it is damaged or not?

• Verdadero
• FALSO

If there are multiple damages in the image, how should the segmentation of the elements, the damages and the annotations be carried out?

• Create an individual segmentation mask for each damage, even if they are of the same type.

  
  

• Create a single segmentation mask that covers all overlapping damages, regardless of the category, and an individual mask for each element.

• Create a single segmentation mask for all overlapping damages of the same type, and individual masks for each element

• Create a single segmentation mask that covers all damages and elements

What is considered a good segmentation mask?

• A mask that tightly covers the entire item or damaged region.

• A mask that covers the entire item or damaged region, regardless of size.

• A mask that covers only the damaged region, leaving out the rest of the item.

• A mask that covers the entire image, including all items and damages.

A link or relationship must always be created to associate the bounding box of the object with the bounding box of the abnormal region

• Verdadero
• FALSO

It is not necessary to create an arrow or relationship to link the bounding box of the object with the bounding box of the abnormal region

• Verdadero
• FALSO