search

chenfengxu714 / SqueezeSegV3

BSD 2-Clause "Simplified" License
220 stars 50 forks source link

Wrong normalization on Kitti? #12

Open autosquid opened 3 years ago

autosquid commented 3 years ago

Around https://github.com/chenfengxu714/SqueezeSegV3/blob/master/src/tasks/semantic/dataset/kitti/parser.py#L171:

proj = torch.cat([proj_range.unsqueeze(0).clone(),
                      proj_xyz.clone().permute(2,0,1),
                      proj_remission.unsqueeze(0).clone()])
proj = (proj - self.sensor_img_means[:, None, None]) / self.sensor_img_stds[:, None, None]

The order of channels in proj is (range, z, x, y, remission), but sensor_image_means is in the order of (range,x,y,z,signal) or something else (based if it's V321 or V353)?

So I guess here's something wrong?

chenfengxu714 commented 3 years ago

Sorry for my late reply, I did't notice the issue recently. Yes, I made a small mistake here, while we also find the normalization on the KITTI doesn't have much influence and just slight drops the performance. It is true that right normalization with accurate mean and std are better. We will update this soon.