Fix implementation of max distance hyperparam for track linking

[shaikh58]

In inference/post_processing.py, function filter_max_center_dist attempts to modify the association matrix during the track assignment step. Its current intent is to disallow track linking if an instance is max_center_dist pixels away from all other instances in the window. However, similar to TrackMate's LAP tracker, a definition that makes more sense is to disallow track linking if an instance is max_center_dist away from the instances in just the previous frame. There are other bugs i.e. with normalizing the distance, and with units; currently distances are in units of fractions of height/width of image; we want the max_center_dist to be in pixels so scale by image height and width

Current:

def filter_max_center_dist(
    asso_output: torch.Tensor,
    max_center_dist: float = 0,
    k_boxes: torch.Tensor | None = None,
    nonk_boxes: torch.Tensor | None = None,
    id_inds: torch.Tensor | None = None,
) -> torch.Tensor:
    """Filter trajectory score by distances between objects across frames.

    Args:
        asso_output: An N_t x N association matrix
        max_center_dist: The euclidean distance threshold between bboxes
        k_boxes: The bounding boxes in the current frame
        nonk_boxes: the boxes not in the current frame
        id_inds: track ids

    Returns:
        An N_t x N association matrix
    """
    if max_center_dist is not None and max_center_dist > 0:
        assert (
            k_boxes is not None and nonk_boxes is not None and id_inds is not None
        ), "Need `k_boxes`, `nonk_boxes`, and `id_ind` to filter by `max_center_dist`"
        k_ct = (k_boxes[:, :, :2] + k_boxes[:, :, 2:]) / 2
        k_s = ((k_boxes[:, :, 2:] - k_boxes[:, :, :2]) ** 2).sum(dim=2)  # n_k

        nonk_ct = (nonk_boxes[:, :, :2] + nonk_boxes[:, :, 2:]) / 2

        dist = ((k_ct[:, None, :, :] - nonk_ct[None, :, :, :]) ** 2).sum(
            dim=-1
        )  # n_k x Np

        norm_dist = dist / (k_s[:, None, :] + 1e-8)
        norm_dist = dist.mean(axis=-1)  # n_k x Np

        valid = norm_dist < max_center_dist  # n_k x Np
        valid_assn = (
            torch.mm(valid.float(), id_inds.to(valid.device))
            .clamp_(max=1.0)
            .long()
            .bool()
        )  # n_k x M
        asso_output_filtered = asso_output.clone()
        asso_output_filtered[~valid_assn] = 0  # n_k x M
        return asso_output_filtered
    else:
        return asso_output```

[shaikh58]

WIP on branch mustafa-fix-max-center-dist