benchmarks.hota.render_pred_panoptic
def render_pred_panoptic(masks: torch.Tensor, categories: torch.Tensor, track_ids: torch.Tensor, height: int, width: int, offset: int, remap: TrackRemap) -> np.ndarrayPaint thing instances as ``semantic * offset + dense_instance``.
Detections with ``track_id < 0`` are skipped (a defensive contract; the
runner only renders kept, tracked instances). Later masks win on overlap
(model panoptic output is non-overlapping in practice).
Source: unitrack/benchmarks/hota/render.py:47