stratomaster31
stratomaster31
I'm working on this model. I've coded the CVAE and I have good results in training phase, but not for test phase... Which are the decoder1 inputs? In the paper...
The zero points in the projection are those zero points that are present in the raw data plus "pure-projection zero points". Latter are pixels in which no raw points fall...
Have you previouslly trained the model and saved any checkpoint? Check if you have changed the path to the checkpoints directory in "eval.py". If no checkpoint exists, I think the...
Another issue is that Table 1 in the paper does not specify which dataset was evaluated (training or evaluation). It should be evaluation, but it is no clear...
I think that the correct weighting follows : if there are 3 classes, and num(class=1) == 3*num(class=2) == 2*num(class=3), then a possible solution is, weights = [1/3 1 2/3]
@BichenWuUCB computing the statistics of your dataset, the inverse ratio of number of pixel's labels I get is 1/15, 1, 59, 62 for 'unknown', 'car', 'pedestrian' an 'cyclist'. I can't...
Well, I didn't dig deeper into it... sorry
What you have to do is to increase the azimuth limits in the 2D projection stage, and it would be advisable to increase the width of the images. In order...
Noticie that AZIMUTH_LEVEL and ZENITH_LEVEL are the sizes of the image, nota the limits of the azimuth and elevación angles
There is a nice quasi-equal implementation in the VoxelNet repository. The difference is that you have to convert the point cloud to spherical coordinates and work only with the azimuth...