search

zhuoxiao-chen / ReDB-DA-3Ddet

[ICCV 2023] Revisiting Domain-Adaptive 3D Object Detection by Reliable, Diverse and Class-balanced Pseudo-Labeling
Apache License 2.0
46 stars 8 forks source link

Waymo → nuScenes #1

Closed AlphaPlusTT closed 1 year ago

AlphaPlusTT commented 1 year ago

The impact of REDB for Waymo adapt to nuScenes is significantly limited, why？

zhuoxiao-chen commented 1 year ago

Thanks for your question.

  1. Challenging Dataset: The nuScene dataset inherently poses significant challenges, resulting in the model (SECOND) achieving fully-supervised training results of only 18.92 (pedestrian) and 11.73 (cyclist) 3D mAP.
  2. Significant domain gap: There exist large domain gaps between Waymo and nuScene, such as very different object densities and beam numbers, which introduces a level of complexity when attempting to adapt to nuScene.

The coexistence of these two factors introduces error accumulation in pseudo-labelling. In the case of ST3D++, which focuses on adapting one model for a single class, the 3D mAP results are just 8.91 (pedestrian) and 4.84 (cyclist). However, when we transition to a more realistic scenario, training all three classes simultaneously while adapting at once, the performance of ST3D++ drops significantly to 1.58 (pedestrian) and 3.74 (cyclist).

Regards, Zhuoxiao (Ivan) Chen