Hi, I would like to get a BEV for each sample (with and without multiple frames) for a defined FOV such as x:(0, 50), y(-25, 25), cell_res=0.1. This should give me a 500x500 BEV as a numpy array which I can later use to train.
Similarly, each object's annotations (center) should also be mapped to corresponding pixels in the BEV.
I have a function of my own to transform the point clouds to flat vehicle and then do this. But, my final evaluation results show something is wrong. I don't get close to 1 AP for detection when I used the GT BEV labels converted to global frame as prediction.
I understand there may be several places where the error could arise from. But I just wanted to know, does the devkit have any such function which gives me a BEV and annotations in the BEV?
I know there is a render_sample_data(), but I think this simply creates an image, but I need it as npy and also need the object centers in the BEV as GT.
Hi, I would like to get a BEV for each sample (with and without multiple frames) for a defined FOV such as x:(0, 50), y(-25, 25), cell_res=0.1. This should give me a 500x500 BEV as a numpy array which I can later use to train. Similarly, each object's annotations (center) should also be mapped to corresponding pixels in the BEV.
I have a function of my own to transform the point clouds to flat vehicle and then do this. But, my final evaluation results show something is wrong. I don't get close to 1 AP for detection when I used the GT BEV labels converted to global frame as prediction.
I understand there may be several places where the error could arise from. But I just wanted to know, does the devkit have any such function which gives me a BEV and annotations in the BEV?
I know there is a render_sample_data(), but I think this simply creates an image, but I need it as npy and also need the object centers in the BEV as GT.
Best Regards Sambit