Sense-X / HoP

[ICCV 2023] Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Ablation study of the bev feature at timestamp t-k

jinhuan-hit opened this issue · comments

Hi, I am wondering do you compare the results of the reconstructed bev feature with the real bev feature at timestamp t-k?

We conducted the experiment on the model below:
Snipaste_2023-08-02_15-17-02

and we found that the performance of HoP branch is:
tmp_hop_his.

Got it! With HoP, the performance of NDS increases to 0.531 from 0.5159.

Hi, @jinhuan-hit . I am afraid that my last answer may mislead you. Let me clarify it.

The performance of HoP branch, which is shown in the second picture, means that it is the detection performance based on reconstructed BEV feature.

The performance increase due to HoP is from 0.513 to 0.531.