Incomplete GT flow

[Shiming94]

Hi Qingwen,

When I visualized the GT flow of some samples, I found some objects still have incomplete flow, as shown in the figure I attached (scene (av2_demo): 25e5c600-36fe-3245-9cc0-40ef91620c22, frame_id:54).

I am sure that I updated the data preprocessing code and re-generated the ground truth with it. Is it still a bug, or is it because of some inherent features hard to fix, like inaccurate bound boxes?

Look forward to your reply! Thank you in advance.

Best,
Shiming

[Kin-Zhang]

Hi shiming,

I think here is the non-ego motion distortion mentioned at HiMo paper; Both are right: 1) the bbx from human annotation is not perfect (i.e., some points belonging to the top of vehicles are missing annotation); 2) undistortion may still miss some points if the velocity annotation from human is not perfect.

OpenSceneFlow/dataprocess/extract_av2.py

Lines 170 to 175 in d5ad83a

	rel_obj_flow = c1_SE3_c0_ego_frame.transform_point_cloud(obj_pts) - obj_pts
	delta_move = abs(np.linalg.norm(rel_obj_flow, axis=0).mean())
	if delta_move > 0.04: # only when it's moving
	c0 = cuboids[0][id]
	c0.length_m += (BOUNDING_BOX_EXPANSION + min(delta_move/2, 2)) # since 180/360 for two LiDARs orientation

The second version ground truth flow is much better than before (v1), but human annotation is never perfect. Factors like speed and insufficient annotation can also lead to the undistortion fix process also. Even with annotations, a complete fix is ​​impossible. I believe this is within the margin of error, and my tests show that DeltaFlow, even when trained on this data, can self-correct such errors. See the image below:(Left: gt; Right: deltaflow-av2 ckpt)

(but the data I show looks like less incomplete flow than your screenshot; I directly downloaded the demo-v2 I uploaded to the Hugging Face: https://huggingface.co/kin-zhang/OpenSceneFlow/blob/main/demo-data-v2.zip)