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

Strangley poor results on pre-trained model #75

Open DaddyWesker opened 2 years ago

DaddyWesker commented 2 years ago

Hello and thanks for sharing your code.

I've got a problem. I've launched train.py --demo on your plane picture and it works fine, but i've tried to use my own pictures and got weirdly poor results. Probably I am missing something. Here is the examples: mustang_re image

plane_new image

I've used png images with black background (first pair), white background and without background (second pair) - same

zergrisk1 commented 2 years ago

Have you removed the background?

DaddyWesker commented 2 years ago

Yes i did. I've downloaded files without background and even tried to create my own no background image. Still the same.

zergrisk1 commented 2 years ago

Try to use the picture same size with the demo-plane picture? Like the file below: 00014

DaddyWesker commented 2 years ago

Your pic works fine. But here is mine

Plane_2

image

kochsebastian commented 2 years ago

@DaddyWesker were you able to resolve the issue? I am facing very similar problems.

DaddyWesker commented 2 years ago

@kochsebastian Unfortunately, no. On some pics model works decent, but on most pics i've used it works as i stated above.