Question: Reference point update strategy

[serend1p1ty]

Great work, congratulations on receiving the best paper award!

I have a question about the reference point update strategy.

velo = output_coords[-1, 0, :, -2:]  # [num_query, 3]
if l2g_r2 is not None:
    # Update ref_pts for next frame considering each agent's velocity
    ref_pts = self.velo_update(
        last_ref_pts[0],
        velo,
        l2g_r1,
        l2g_t1,
        l2g_r2,
        l2g_t2,
        time_delta=time_delta,
    )
else:
    ref_pts = last_ref_pts[0]

dim = track_instances.query.shape[-1]
track_instances.ref_pts = self.reference_points(track_instances.query[..., :dim//2])
track_instances.ref_pts[...,:2] = ref_pts[...,:2]

Why do we need to update the z-coordinate of the reference point with self.reference_points after updating reference point with velocity?

[serend1p1ty]

@YTEP-ZHI Hello, maybe we can remove these redundant codes?

dim = track_instances.query.shape[-1]
track_instances.ref_pts = self.reference_points(track_instances.query[..., :dim//2])
track_instances.ref_pts[...,:2] = ref_pts[...,:2]