search

KTH-RPL / DeFlow

[ICRA'24] DeFlow: Decoder of Scene Flow Network in Autonomous Driving
BSD 3-Clause "New" or "Revised" License
102 stars 4 forks source link

argoverse2 val experiment #8

Closed wangyunlhr closed 5 days ago

wangyunlhr commented 6 days ago

Hi, Thank you for sharing this great project! I’m currently trying to load the model(deflow_best.ckpt) and test it on the argoverse2 validation dataset, but I’m noticing that the results are slightly worse than expected. Dataprepocess: I didn't use the last line to transform. [ego2sensor_pose] 屏幕截图 2024-10-30 195021 Could it be an issue with the data processing? I’m wondering about the coordinate transformation, as I’m not quite sure why it’s necessary.

2

Thank you for your help!

Kin-Zhang commented 6 days ago

Yes, I think you should follow the dataprocess correctly and include that part, since it is trained on the sensor frame point cloud, the different is shown here: (if you comment the code will be the left one, and the correct is right one)

And this result is I run just now using this repo with the deflow_best weight: image

wangyunlhr commented 6 days ago

Thank you for your help! After transforming the points to the sensor coordinate system, should the flow labels also be transformed to the sensor coordinate system?

Kin-Zhang commented 5 days ago

No need, since this sensor coordinate is just higher the z a little bit, and the flow label is calculated the relative flow between p0 and p1, so it won't be affect is p0 p1 both transform to the same coor.

wangyunlhr commented 5 days ago

Got it, thank you very much!