search

ThibaultGROUEIX / AtlasNet

This repository contains the source codes for the paper "AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation ". The network is able to synthesize a mesh (point cloud + connectivity) from a low-resolution point cloud, or from an image.
http://imagine.enpc.fr/~groueixt/atlasnet/
MIT License
677 stars 119 forks source link

How to train a SVR(nb_primitives:1) from scratch? #68

Closed sumai-1998 closed 3 years ago

sumai-1998 commented 3 years ago

You used a pretrained model in AtlasNet-master/training/trained_models/atlasnet_singleview_1_sphere called "log/atlasnet_singleview_1_sphere_tmp_True/network.pth", which has a "-5.5" train/val loss in 'Curves in log'.
However if i train without a pretrained model, i just reach at '-5 train and -2 val' loss, which performs worse in 'display' and some airplane, and lack details in the other calsses.

sumai-1998 commented 3 years ago

How to first train and get a log/atlasnet_singleview_1_sphere_tmp_True/network.pth? ：-）

sumai-1998 commented 3 years ago

This is mine train from scratch with the default settings, in SVR_1 Default curve This is your in training/trained_models/atlasnet_singleview_1_sphere Authors curve

ThibaultGROUEIX commented 3 years ago

Hi @4ssass1n0r, Single-view networks are trained in 3 steps: 1- Train the atlasnet decoder in an autoencoder setting (with a pointnet encoder) 2- train the resnet encoder with the freezed atlasnet decoder 3- fine-tune all weights of the SVR (resnet+atlasnet) You can find more details in the paper. Best regards, Thibault