Extract mapping of original to tracked cell labels

[hzwirnmann]

Dear all,

Thanks for publishing and documenting your package in such a great way.

I have one further question:

Is there a proper way to extract the mapping of old to new cell labels? E.g. a DataFrame with three columns "time step", "original label", "new label". I would need this data to correct entries in a database that were stored at creation time of the mask. So far, I'm comparing all values of the original and ctc_masks (yes, quite a lot of comparisons) to create a mapping dict. Is there a more elegant solution to this problem?

Thank you in advance.

Best wishes
Henning

[C-Achard]

Dear Henning,

Hello, I may be able to help as I had to do something similar some time ago.

Please find below a code snippet generated from my main code, it may not work exactly as-is, but it should get you most of the way there (if that solution happens to be suitable for your use case anyway).

The way I did this was to use traccuracy to match labels; see below.
Also, you can use any arbitrary pair of labels, no need for GT, it should still attempt to match the two sets no matter what.
(Perhaps you can even omit metrics and only run the matching)

import gzip
import pickle
from pathlib import Path
import pandas as pd
from traccuracy import run_metrics
from traccuracy.loaders import load_ctc_data
from traccuracy.matchers import CTCMatcher
from traccuracy.metrics import CTCMetrics

def run_tracking_and_extract_matching(img_path, det_path, model_dir, outdir):
    """
    Run tracking and extract the matching between original labels and tracked labels.

    Args:
        img_path (Path): Path to the image folder.
        det_path (Path): Path to the ground truth folder.
        model_dir (Path): Path to the model folder.
        outdir (Path): Output directory to save results.

    Returns:
        pd.DataFrame: DataFrame with GT to predicted label mapping.
    """
    # Ensure output directory exists
    tracked_path = outdir / "tracked"
    tracked_path.mkdir(parents=True, exist_ok=True)

    # Run tracking if predictions do not exist
    # from Trackastra.track_from_disk or other

    # Load ground truth and predictions
    tracked = load_ctc_data(tracked_path, tracked_path / "man_track.txt", name="Tracked")
    gt = load_ctc_data(det_path, det_path / "man_track.txt", name="GT")

    # Run metrics and get matched graphs
    _, matched_graphs = run_metrics(
        gt_data=gt,
        pred_data=tracked,
        matcher=CTCMatcher(),
        metrics=[CTCMetrics()],
    )

    # Extract label matching
    gt_to_pred = matched_graphs.gt_pred_map
    gt_pred_df = pd.DataFrame([
        {"gt_label": gt_label, "pred_label": pred_label}
        for gt_label, pred_label in gt_to_pred.items()
    ])
    return gt_pred_df

if __name__ == "__main__":
    # Example usage
    img_path = Path("path/to/images")
    det_path = Path("path/to/ground_truth")
    model_dir = Path("path/to/model")
    outdir = Path("path/to/output")

    label_matching_df = run_tracking_and_extract_matching(img_path, det_path, model_dir, outdir)
    print(label_matching_df.head())

matched_graphs.gt_pred_map should contain what you need for downstream processing.

Hope this helps,
Cyril

[hzwirnmann]

Hi Cyril,

Thank you for the quick reply.

I went through your code and also through the documentation of traccuracy, but I wonder how I can tailor it to my needs. For this in order to work, I would need both the old labels and the new labels as TrackingGraph. Would you have any idea how to do that without saving files to disk?
You're using load_ctc_data, which does read from disk. As an alternative, I see that there is load_point_data in traccuracy that also accepts a DataFrame. However, it expects complete data with coordinates. In Trackastra, neither track_graph nor ctc_tracks seem to match the expected format.

Best wishes
Henning