Issue with AWS Rekognition QualityFilter not Filtering Low-Quality Faces

Question

We are only adding new users into our collection when IndexFaces QualityFilter passing 'HIGH' threshold. However, we noticed recently that low quality photos are also getting accepted by this level of filtration.

For example, from the output below, the photo (and face) quality is so low, even 'Sharpness' equals to 3. But despite that, the face on the photo is passed through and accepted by QualityFilter High.

```
[cloudshell-user@ip-10-130-84-170 ~]$  aws rekognition index-faces --collection-id "test_01" --quality-filter "HIGH" --detection-attributes "ALL" --image "S3Object={Bucket="",Name=".jpg"}" --max-faces 1

"Quality": {
                    "Brightness": 74.86048126220703,
                    "Sharpness": 3.3018569946289062 <--------------------------
                },
                "Confidence": 99.87649536132812,
                "FaceOccluded": {
                    "Value": false,
                    "Confidence": 87.64056396484375
                },
                "EyeDirection": {
                    "Yaw": -6.743906021118164,
                    "Pitch": -12.818127632141113,
                    "Confidence": 9.133011189987883e-06
                }
            }
        }
    ],
    "FaceModelVersion": "6.0",
    "UnindexedFaces": []
}
```

We have conducted various tests on FaceModelVersion 6.0 (eu-central-1) and 7.0 (eu-west-1), but the results are the same - QualityFilter is not working correctly.

Why QualityFilter input parameter is not filtering out detected faces, even if they don’t meet a required quality bar?

Regards,
Almaz

Answer

It is possible for a face to pass the QualityFilter threshold of 'HIGH' but still appear low quality. The QualityFilter evaluates two factors - sharpness and brightness. As long as the face meets or exceeds the thresholds for those factors, it will pass.

These may be the reasons:
* The sharpness of the particular face was rated a 3, which may have met the internal threshold even though it looks blurry to a human. Sharpness is evaluated on a scale of 0 to 1.
* Other factors like pose, size or lighting could have negatively impacted appearance but not affected the sharpness/brightness metrics enough to fail the filter.
* The thresholds that determine 'HIGH' quality may need to be adjusted higher for your specific use case and image characteristics.

To ensure only high quality faces are indexed, some options are:
* Manually review faces added to verify quality meets needs
* Experiment with increasing the QualityFilter threshold
* Add multiple reference images per person in various conditions to improve matching ability despite some lower quality samples

Issue with AWS Rekognition QualityFilter not Filtering Low-Quality Faces

관련 콘텐츠