delta_t argument ignored in prediction stage of Trackastra.track()?

[aarpon]

From our experiments, it looks that delta_t has little if no effect on the end result of tracking.

Indeed, trackastra.model.model_api.Trackastra.track() calls trackastra.model.model_api.Trackastra._predict() without forwarding **kwargs, as follows:

        predictions = self._predict(
            imgs,
            masks,
            normalize_imgs=normalize_imgs,
            progbar_class=progbar_class,
            n_workers=n_workers,
            batch_size=batch_size or self.batch_size,
        )

The _predict() method, in turns, calls

        predictions = predict_windows(
            windows=windows,
            features=features,
            model=self.transformer,
            edge_threshold=edge_threshold,
            spatial_dim=masks.ndim - 1,
            progbar_class=progbar_class,
            batch_size=batch_size,
        )

The method predict_windows() takes delta_t as input argument, but it defaults to 1:

def predict_windows(
    windows: list[dict],
    # features: list[WRFeatures],
    # model: TrackingTransformer,
    features: list,
    model,
    intra_window_weight: float = 0,
    delta_t: int = 1,
    edge_threshold: float = 0.05,
    spatial_dim: int = 3,
    batch_size: int = 1,
    progbar_class=tqdm,
) -> dict:
...

In contrast, delta_t seems to be passed to trackastra.model.model_api.Trackastra._track_from_predictions() via **kwargs:

track_graph = self._track_from_predictions(predictions, mode=mode, **kwargs)

So, is it correct that delta_t seems to be ignored in the prediction stage and only be used in building the graph for the solver? If this is true, this seems to strongly limit (or even nullify) the effect of delta_t > 1 in graph construction, since the prediction path appears to generate associations only up to delta_t = 1.

[bentaculum]

Our released pre-trained models only handle delta_t = 1 indeed, having it configurable is legacy only.

[aarpon]

Ah, I see. I then misunderstood the link with the window parameter in the training: I thought that if the model is trained with window = 4 (for example), then delta_t could also vary from 0 to 3 (beyond that having no effect).