calibration: use .iloc[0] so positional access works under pandas 2.x

[narr2atnarrable]

Summary

Three calibration modules read the first row of their per-image metadata DataFrame with label-based lookup:

• sn_gamestate/calibration/baseline.py — metadatas["lines"][0]
• sn_gamestate/calibration/nbjw_calib.py — metadatas["keypoints"][0]
• sn_gamestate/calibration/pnlcalib.py — metadatas["keypoints"][0] and metadatas["lines_det"][0]

Under pandas 2.x these crash as soon as the DataFrame's index is not 0-based. The ExternalVideo wrapper (dataset=video) keeps the original image_id as the index, so the first frame immediately raises:

File ".../sn_gamestate/calibration/nbjw_calib.py", line 179, in process
    predictions = metadatas["keypoints"][0]
  File ".../pandas/core/series.py", line 1121, in __getitem__
    return self._get_value(key)
  File ".../pandas/core/series.py", line 1237, in _get_value
    loc = self.index.get_loc(label)
  File ".../pandas/core/indexes/range.py", line 417, in get_loc
    raise KeyError(key)
KeyError: 0

.iloc[0] selects the first positional row regardless of the index, which is what all three modules intend.

Reproduction

Any pandas 2.x install, dataset=video (ExternalVideo wrapper or any path that preserves the caller's index), calibration module set to baseline, nbjw_calib or pnlcalib:

pip install "pandas>=2.0"
python -m tracklab.main -cn soccernet \
    dataset=video \
    modules.calibration=nbjw_calib \
    dataset.video_path=/path/to/clip.mp4

First frame raises KeyError: 0 from the calibration module.

Change

Four label-lookups → positional lookups:

 # sn_gamestate/calibration/baseline.py
-        predictions = metadatas["lines"][0]
+        predictions = metadatas["lines"].iloc[0]

 # sn_gamestate/calibration/nbjw_calib.py
-        predictions = metadatas["keypoints"][0]
+        predictions = metadatas["keypoints"].iloc[0]

 # sn_gamestate/calibration/pnlcalib.py
-        keypoints = metadatas["keypoints"][0]
-        lines = metadatas["lines_det"][0]
+        keypoints = metadatas["keypoints"].iloc[0]
+        lines = metadatas["lines_det"].iloc[0]

Risk

None on datasets that do reset_index(drop=True) before the calibration module runs — then index 0 == position 0 and .iloc[0] returns the same row as [0] would. The change is strictly a bug fix: it makes the code work for the full range of DataFrame indexes, not just the 0-based case.

Context

Encountered running sn-gamestate end-to-end on a Blackwell (sm_120) GPU with a fresh pandas 2.x install. Filed alongside a sibling PR to TrackingLaboratory/tracklab for a related Windows-path fix; both were needed to get a mixed-OS pipeline running.