[BUG] Failed to populate metadata on 72 samples. Use dataset.exists("metadata", False) to retrieve them

[harry-esmart]

Does anyone know why I faced this error: [BUG] Failed to populate metadata on 72 samples. Use dataset.exists("metadata", False) to retrieve them while calculating evaluate_detections?

Computing metadata...
100% |█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 72/72 [791.6ms elapsed, 0s remaining, 91.0 samples/s]
Failed to populate metadata on 72 samples. Use dataset.exists("metadata", False) to retrieve them
{
'samples_count': 72,
'samples_bytes': 510162,
'samples_size': '498.2KB',
'media_bytes': 0,
'media_size': '0.0B',
'total_bytes': 510162,
'total_size': '498.2KB',
}
Failed to populate metadata on 72 samples. Use dataset.exists("metadata", False) to retrieve them
Threshold: 0.5
Evaluating detections...
100% |█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 72/72 [1.8s elapsed, 0s remaining, 40.7 samples/s]
precision recall f1-score support

 flashed       0.00      1.00      0.00         1

contaminated 0.00 1.00 0.00 1

micro avg 0.00 1.00 0.00 2
macro avg 0.00 1.00 0.00 2
weighted avg 0.00 1.00 0.00 2

[harry-esmart]

Another question is that this problem also causes 0 in precision.

[brimoor]

Have you inspected view = dataset.exists("metadata", False) to see what the problem is? For example, the image paths could be nonexistent.

This error is coming from dataset.compute_metadata(), which is internally used to populate things like image dimensions, which are necessary for certain computations. You could try directly running that a second time to see if those images are still unable to be processed.

[harry-esmart]

Yeah, the image paths do not exist but that should not be affecting the evaluation computations right? As it only needs bboxes from ground-truth and prediction.

[brimoor]

The evaluation protocols whatever method you're calling needs information about absolute image dimensions for certain things. For example, labels themselves are stored in FO as relative values, so the metadata field either needs to be pre-populated with ImageMetadata instances with their height and width populated, or the images need to exist on disk so that dataset.compute_metadata() can populate this data on-the-fly.

[harry-esmart]

@brimoor So what you meant is that image paths are a prerequisite to calculating precision|recall|f1 correctly?

[brimoor]

Actually I'm confused. What code are you running exactly? This works perfectly fine with non-existent filepaths:

import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

dataset.set_values("filepath", ["bad.jpg"] * len(dataset))

results = dataset.evaluate_detections(
    "predictions",
    gt_field="ground_truth",
    eval_key="eval",
    compute_mAP=True,
)

[harry-esmart]

Hi @brimoor,
It is true that the code can be run with non-existent filepaths. But during its running, there is an warning/error saying Failed to populate metadata on 72 samples. Use dataset.exists("metadata", False) to retrieve them. And the calculated metric results are not correct, precision/recall/f1. I am not sure it is because of the warning/error above or something else that causes the problem.