Hungarian matching fails with infeasible cost matrix when scores are all NaN

[talmo]

Bug

hungarian_matching() in sleap_nn/tracking/utils.py crashes with ValueError: cost matrix is infeasible when the cost matrix passed to scipy.optimize.linear_sum_assignment() contains only inf values.

How to reproduce

Run tracking with --track_matching_method hungarian on data where some frames produce all-NaN scores (e.g., when no candidates pass the min_match_points filter or when keypoints contain NaN values):

sleap-nn track \
    -i predictions.slp \
    -o tracked.slp \
    --tracking \
    --track_matching_method hungarian \
    --scoring_method euclidean_dist \
    --features keypoints \
    --candidates_method local_queues \
    --max_tracks 2

Root cause

In tracker.py, scores_to_cost_matrix() converts NaN scores to inf costs:

def scores_to_cost_matrix(self, scores: np.ndarray):
    cost_matrix = -scores
    cost_matrix[np.isnan(cost_matrix)] = np.inf  # NaN → inf
    return cost_matrix

When all scores for a frame are NaN (e.g., empty scores_trackid list → np.nanmean([]) returns NaN), the entire cost matrix becomes inf, which linear_sum_assignment() considers infeasible.

greedy_matching() does not have this problem because np.argsort() silently handles inf values by sorting them to the end.

The current hungarian_matching() has no such handling:

def hungarian_matching(cost_matrix: np.ndarray):
    row_ids, col_ids = linear_sum_assignment(cost_matrix)  # fails on all-inf matrix
    return row_ids, col_ids

Suggested fix

Replace inf/nan values with a large finite number before calling linear_sum_assignment:

def hungarian_matching(cost_matrix: np.ndarray):
    cost_matrix = np.copy(cost_matrix)
    bad = ~np.isfinite(cost_matrix)
    if bad.any():
        finite_vals = cost_matrix[~bad]
        fill = (np.abs(finite_vals).max() * 10 + 1) if finite_vals.size > 0 else 1e6
        cost_matrix[bad] = fill
    row_ids, col_ids = linear_sum_assignment(cost_matrix)
    return row_ids, col_ids

This makes the assignment feasible while still strongly penalizing unmatched pairs, preserving the intended behavior.

Environment

• sleap-nn 0.1.0 (installed from source, main branch)
• scipy 1.17.1
• Python 3.12