Giters

microsoft / SpareNet

Style-based Point Generator with Adversarial Rendering for Point Cloud Completion (CVPR 2021)

Home Page:https://arxiv.org/abs/2103.02535

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Unable to Reproduce Results

chekirou opened this issue · comments

commented

Hi,
Great paper ! I am trying to recreate your results using selected chunks from the ScanNet dataset. I'm unable to reproduce your results using the pretrained sparenet_model for shape categories that were included in the training data. I processed the points as follows :

diag = points.max(dim=1).values - points.min(dim=1).values # shape :  1 x N x 3
norm = 1 / torch.linalg.norm(diag)
c = points.mean(dim=1)
points = (points - c) * norm

it works fine for points from the val or train set from shapenet but not for clouds such as this one :
input_chair
I sampled ~1000 points, it always gives this kind of reconstruction for the coarse point cloud:
Capture_res_chair
Is there some preprocessing step I'm missing ?
Thank you in advance.

commented

Thanks for the question! We haven't incorporated the ScanNet dataset. You can refer to the dataloader of ShapeNet and adapt it for ScanNet:

SpareNet/datasets/data_loaders.py

Line 145 in 8bbc200

transforms = self._get_transforms(self.cfg, subset)