Need one image to have multiple labels

Question

Hi, I am looking for a way to create datasets using manifest or auto labeling feature or any other way to allow me to label multiple objects in an image so that each object I select has a label.   
  
Tried to create datasets using manifest so that the manifest only has class_id and class-map attributes but not having any boundingbox attributes. I was able to create such datasets. From the dataset GUI, I can see each each image has multiple labels associated. However training fails with error "The manifest file contains too many invalid data objects."  
  
I made another attempt by putting same image that has multiple objects in multiple folders that represent their labels. I was able to create datasets and even successfully trained that dataset-pair. But  from viewing the test result, no single image correctly labeled twice.  
  
Do you support my scenario?  
  
thank you.

Answer

That manifest format works. Thank you.

Answer

Hey mymingle  
  
Unfortunately the custom labels console does not support attaching multiple labels to a single image using the autolabel feature. I have passed on this feedback to the corresponding team.  
  
You can follow the below sample format when manually creating a multi label manifest file (expanding on the example specified here https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/cd-required-fields.html):  
  
 {  
    "source-ref": "S3 bucket location", # Required  
    "sport0":0, # FIRST annotation  
    "sport0-metadata": { # Required  
        "class-name": "football", # Required  
        "confidence": 0.8, # Required  
        "type":"groundtruth/image-classification", # Required  
        "job-name": "identify-sport", # Not required  
        "human-annotated": "yes", # Required  
        "creation-date": "2018-10-18T22:18:13.527256" # Required  
    },  
    "sport1":1, # SECOND annotation  
    "sport1-metadata": { # Required  
        "class-name": "ball", # Required  
        "confidence": 0.8, # Required  
        "type":"groundtruth/image-classification", # Required  
        "job-name": "identify-sport", # Not required  
        "human-annotated": "yes", # Required  
        "creation-date": "2018-10-18T22:18:13.527256" # Required  
    }  
}  # end of annotations for 1 image   
  
The above line attaches labels "football" and "ball" to the same image.  
  
You can use any string instead of "sport0", "sport1" etc but you must ensure there is a corresponding "-metadata" field. For e.g. if you use "some-label1", ensure it is followed by a "some-label1-metadata" section.  
  
Each "some-label", -metadata" pair represents ONE annotation.  
  
Hope this helps.

Answer

Reposting example with formatting -

```
 {
    "source-ref": "S3 bucket location", # Required
    "sport0":0, # FIRST annotation
    "sport0-metadata": { # Required
        "class-name": "football", # Required
        "confidence": 0.8, # Required
        "type":"groundtruth/image-classification", # Required
        "job-name": "identify-sport", # Not required
        "human-annotated": "yes", # Required
        "creation-date": "2018-10-18T22:18:13.527256" # Required
    },
    "sport1":1, # SECOND annotation
    "sport1-metadata": { # Required
        "class-name": "ball", # Required
        "confidence": 0.8, # Required
        "type":"groundtruth/image-classification", # Required
        "job-name": "identify-sport", # Not required
        "human-annotated": "yes", # Required
        "creation-date": "2018-10-18T22:18:13.527256" # Required
    }
}  # end of annotations for 1 image 
```

Edited by: awsrakesh on Mar 24, 2020 1:56 PM

Answer

Hi, Could I get some expert opinions for my questions? thx.

Need one image to have multiple labels

Relevant content