Custom metrics that depends on image_path?

[Anton-Cherepkov]

💡 Your Question

Hello, community

Problem description:
I want to create a custom class PrecisionRecallROIMetrics(DetectionMetrics).

The class PrecisionRecallMetrics will contain logic for calculating detection metrics (e.g. precision, recall) for some Region of Interest (arbitrary region of image defined by polygon). The Region of Interest will be unique for each image in the dataset and can be somehow retrieved by the path to the image.

My current problem with implementing such a class is the following:

• the update method of DetectionMetrics has fixed inputs which do not include the img_paths:

def update(
  self,
  preds,
  target: torch.Tensor,
  device: str,
  inputs: torch.tensor,
  crowd_targets: Optional[torch.Tensor] = None) -> None:

Question:
Is there any easy way to populate the img_path to the DetectionMetrics object?

Versions

No response