search

HuangJunJie2017 / BEVDet

Code base of the BEVDet series .
Apache License 2.0
1.45k stars 266 forks source link

Question about the way to build cost volume #278

Closed Yzichen closed 1 year ago

Yzichen commented 1 year ago

BEVStereo uses the difference sum of the corresponding features to represent the matching cost when building the cost volume. Is this better than using the dot product between features? Besides, Which paper did you reference for your implementation?

Yzichen commented 1 year ago

In addition, this implementation only utilises the temporal information from the same camera to compose stereo vision, and does not utilise the stereo information across cameras. Do I understand you correctly?

HuangJunJie2017 commented 1 year ago

@Yzichen Just a private implementation~ in my ablation L1 norm is better than dot product

HuangJunJie2017 commented 1 year ago

@Yzichen yes, I haven't used the stereo information across cameras as it doesn't contribute to the performance. maybe my implementation in this direction has some mistakes. you can try it on your own.

Yzichen commented 1 year ago

@HuangJunJie2017 Thank you for your prompt reply and for your work.