Skip to content

Latest commit

 

History

History
100 lines (77 loc) · 4.38 KB

README.md

File metadata and controls

100 lines (77 loc) · 4.38 KB

Annotation Format

The format of provided annotation files.

Annotation Diagram

Image-level Part State Annotations for HICO

Format Details

The annotation results have the same format as follows:

{  
    ...
    "HICO_train2015_000xxxxx.jpg":{  
        "hoi_id":[490, 492, 493, 494],
        "parts_list":[0, 0, 0, 0, 0, 0, 1, 0, 0, 0],
        "foot_list":[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1],
        "leg_list":[0, 0, 0, 0, 0, 0, 0, 0, 0, 1],
        "arm_list":[0, 0, 0, 0, 1],
        "hand_list":[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        "hip_list":[0, 0, 0, 0, 1],
        "head_list":[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]
    },
    "HICO_train2015_000xxxxx.jpg": {
        ...
    }
    ...
}

In JSON format, where key is the name of the image, and the values include:

  • hoi_id: Image-level HOI labels(same setting as HICO, txt format)
  • parts_list: Part-State binary labels of length 10, where 1 indicates that the corresponding human-part is active in the image. The human-part setting is same as Pair-wise. The order is Right_foot, Right_leg, Left_leg, Left_foot, Hip, Head, Right_hand, Right_arm, Left_arm, Left_hand. You can refer to this image for visualization.
  • foot_list, leg_list, arm_list, hand_list, hip_list, head_list: Part-State labels of 6 combined human-parts(merge left and right in foot, leg, arm, hand by OR operation). They are represented in one-hot way, and the value of last index is 1 means that this human-part is inactive in this image. Please refer to our Part-State setting for more details.

Instance-level Part State Annotations for HICO-DET

Different with image-level, instance-level part states are designed for each active person. That is, each GT human instance has body part state labels according to the actions.

Format Details

The annotation results have the same format as follows:

...
{
	u'dataset': u'hico-det',
	u'labels': [
	{
	  	u'action_labels': [
			{u'human_part': 0, u'partstate': 0},
			{u'human_part': 3, u'partstate': 0}
		],
		u'height': 458,
		u'hoi_id': 524,
   		u'human_bbox': [1, 60, 395, 399],
   		u'object_bbox': [88, 247, 637, 449],
   		u'width': 640
   	},
  	{
  		u'action_labels': [
  			{u'human_part': 0, u'partstate': 0},
		   {u'human_part': 3, u'partstate': 0}
		],
	   u'height': 458,
   		u'hoi_id': 525,
   		u'human_bbox': [2, 54, 397, 404],
   		u'object_bbox': [170, 292, 637, 452],
   		u'width': 640
   	}
   	...
   	],
 	u'path_prefix': u'hico_20160224_det/images/train2015'
 }
 ...

In JSON format, where key is the name of the image, and the values include:

  • action_labels: For each human-object pair, we provide the part state labels in human_part and partstate pairs. The human\part order is Right_foot, Right_leg, Left_leg, Left_foot, Hip, Head, Right_hand, Right_arm, Left_arm, Left_hand. You can refer to this image for visualization. Partstate label lists are provided in our paper.
  • hoi_id: Image-level HOI labels(same setting as HICO, txt format)

HAKE-Large Part State Annotations

Format Details

The annotation results share the same format with Instance-level Part State Annotations for HICO-DET.

About the part-state setting

We adopt different part-state settings for image-level, instance-level tasks. You can refer to the following figure for visualization.