The experimental comparison with BEVformor is mentioned in this paper. Is it mean that to retrained BEVformer network after converting the four-way fisheye image into a dedistorted image? Whether the four-way dedistorted image can extract BEV features effectively。
thanks for sharing this project, i also have real world fisheye images to predict but got no GT, may i ask when will you release the style transfer tools and process?
as mentioned in the paper,
We run neural style transfer [41] on the simulated FB-SSEM dataset images using samples of realworld images as style images. Additionally, we note that the simulated fisheye cameras differ from those mounted on the actual car in terms of the valid regions of the image (i.e., regions not covered by the camera or the ego car). We account for this difference during training by predefining the valid region for every camera in the training data to be the same as that of the corresponding real-world camera.
I would like to use FB-SSEM dataset for a task that requires BEV rgb images but these images are missing for almost all sequences (except Images0 and Images1). Will the remaining BEV images be released ?
Thank you in advance for your answer and sharing this amazing work !