search

autonomousvision / occupancy_networks

This repository contains the code for the paper "Occupancy Networks - Learning 3D Reconstruction in Function Space"
https://avg.is.tuebingen.mpg.de/publications/occupancy-networks
MIT License
1.5k stars 292 forks source link

Results not consistent with paper #24

Closed rozentill closed 4 years ago

rozentill commented 4 years ago

Hi there,

I just tried test your model on single view image reconstruction task, i.e. configs/img/onet_pretrained.yaml .

However, I noticed two results which are not consistent with the paper, actually worse than those. Did I use some wrong reconstruction methods or the model was retrained after the publication? If so, any idea on how to improve the quality?

Input 1: 00_in

Output: Capture

Paper: Capture

Input 2: 07_in

Output: Capture

Paper: Capture

Thanks, Ryan

LMescheder commented 4 years ago

Hi @rozentill, please see the notes in the Readme:

In our paper we used random crops and scaling to augment the input images. However, we later found that this image augmentation decreases performance on the ShapeNet test set. The pretrained model that is loaded in configs/img/onet_pretrained.yaml was hence trained without data augmentation and has slightly better performance than the model from the paper.

Indeed, when I change

inherit_from: configs/img/onet_pretrained.yaml

to

inherit_from: configs/img/onet_legacy_pretrained.yaml

in configs/demo.yaml, I get image image

which should be close to what we presented in the paper. There might be some remaining differences (maybe the checkpoint is from a slightly different epoch, I am not sure anymore), but they should be rather small. Also note that quantitatively, when averaged over all models, configs/img/onet_pretrained.yaml achieves slightly better performance than the old version.

rozentill commented 4 years ago

Thank you for your reply. This solved my problem.