otakuxiang/circle

The repository contains the implementation of our ECCV 2022 paper: CIRCLE: Convolutional Implicit Reconstruction and Completion for Large-scale Indoor Scene.

CIRCLE

This repository contains the implementation of our ECCV 2022 paper: CIRCLE: Convolutional Implicit Reconstruction and Completion for Large-scale Indoor Scene.

Haoxiang Chen, Jiahui Huang, Tai-Jiang Mu, Shi-Min Hu

Introduction

CIRCLE is a framework for large-scale scene completion and geometric refinement based on local implicit signed distance function. Please refer to our technical report for more details.

Implementations

We now provide implementation based on Pytorch and we will soon provide jittor version which is faster for training. Please refer to the corresponding folders jittor/ and pytorch/ for specific build and running instructions.

circle's Issues

mesh reconstruction has patterns.

Hello again.
I tested the network with your pretrained models by running run_matterport.py with these settings:
checkpoint = "pretrain" # experiment_name, sys.argv[1]
c_num = 100000 # yours is c_num = 100000.
model, args_model = net_util.load_origin_unet_model(f"pre-trained-weight/{checkpoint}/hyper_small.json",c_num,use_nerf= False,layer = 5)
..............
..............
data_dir = "/media/data2/data/Matterport3d/datav1/matterport_noise_result/" # contains noisy point cloud.
scene = "2t7WUuJeko7"
region = "region0"#

I tested with your pretrained model. the mesh results look like the following:

Do you know why it contains the patterns? Do you have any idea how to remove it ?
Thanks a lot for your time !
Regards

dataset folder

Hello, I would like to reproduce your code and find that your dataset folder is particularly messy. I don't know the specific meaning. The specific documents are as follows

sample_ matterport.py

src_ root = f"/home/chx/data_disk/MatterPort3D/"

tar_ root = f"/home/chx/nas/disk_0/matterport_result/{sys.argv[1]}/"

add_ noise.py

result_ root = "/home/chx/ssd/chx_data/MatterPort/"

crop_ small.py

result_ root = f"/home/chx/data_disk/MatterPort3D/small_{rname}_{nname}/"

train.py

f = open("/home/chx/data_disk/MatterPort3D/scenes_train.txt")

train_ dataset = NoisyOtherMatterPortDataset("/home/chx/ssd/MatterPort3d/",scenes,expand=True,voxel_size = voxel_size,snum = snum,layer = layer)

run_ matterport.py

data_ dir = "/home/chx/nas/other_disk/CVPR2022CIRCLE/noisy_depth/"

result_ root = "/home/chx/nas/disk_0/matterport_result/test"

I didn't quite understand this one at the end

EXPERIMENT_ NAME> <CHECKPOINT_ ITER> <MATTERPORT_ SCENE> <MATTERPORT_ What does it represent

python run_ matterport.py <EXPERIMENT_ NAME> <CHECKPOINT_ ITER> <MATTERPORT_ SCENE> <MATTERPORT_ REGION>

Looking forward to your answer

Environment configuration

Hi! Thank you for your work!
Could you please share the requirements for the environment? Which cuda do you have and which versions of torch, open3d and other libraries do you use?