Open Yeoninm opened 1 year ago
You can refer to our TABLE 3. Bevpool can explicitly utilize the depth information and significantly boost the performance.
Why does MapTRv2 only have BEVEncoder based on bevpool? Have you tried BEVEncoder based on GKT
what about the trade-off of accuracy and time-consumption on gpu?
I see that you have also BEVFormerEncoderDepth
, is it a combination of depth on top of bevencoder?
Do you have any experimental results on this?
Why does MapTRv2 only have BEVEncoder based on bevpool? Have you tried BEVEncoder based on GKT