Resolve residual noise in the PointOdyssey dataset

Resolve the remaining noisy point annotations in the corrected PointOdyssey dataset that the authors were unable to fix, so that the dataset provides reliable supervision for training and evaluating long-horizon 3D point trackers.

Background

The paper describes a programmatic correction pipeline for several PointOdyssey annotation problems, including off-by-one camera extrinsics, corrupted visibility flags, and a temporally incoherent sequence. The authors report that these corrections reduce loss spikes during training and that the affected sequence is excluded from the training and evaluation splits.

Despite these corrections, the dataset still contains noisy point annotations. This unresolved data-quality issue may impair supervision and evaluation for TrackEverything and other 3D point-tracking methods, motivating further identification and correction of the remaining errors.