Alex-fishred
commented
10 months ago Sorry, I may not fully understand what you mean Now my requirement is to put the road video I recorded in advance (only the front view) into the BEVFormer model to generate a 3D frame and road segmentation From your answer, I understand that there is a chance that inputting only the front view will work, but I don't quite understand what you mean by "This result is obtained after training on the nusence dataset. "
My mother tongue is Chinese If your native language is also Chinese, you can answer in Chinese and maybe I will understand you better.
Bin-ze
This result is obtained after training on the nusence dataset. The implementation logic of the segmentation part is to add a segmentation branch after the BEV feature map and use segmentation annotations centered on the self-vehicle within the perception range for supervision.
If you use a single image such as the front view, it may take some effort to work, but from the perspective of the BEVformer algorithm, I think it is possible. It seems that it can be achieved by just modifying some configurations, but the effect is difficult to evaluate without conducting experiments.