search

QianyiWu / objectsdf_plus

:first_quarter_moon: [ICCV'23] Pytorch implementation of "ObjectSDF++: Improved Object-Compositional Neural Implicit Surfaces"
https://qianyiwu.github.io/objectsdf++
MIT License
143 stars 5 forks source link

May I know the scene id used in ScanNet? #9

Open Pixie8888 opened 1 year ago

Pixie8888 commented 1 year ago

Hi,

May I know the scene ids used in ScanNet? I want to fairly compare objectsdf++ with rico, but it seems that both use different scenes in ScanNet.

Also, may I know how to pre-process ScanNet? Particularly, how to compute "instance_mapping.txt", "cameras.npz", depth and normal maps?

QianyiWu commented 1 year ago

You might refer to this script. The scan ID list is given at here. Remember to modify the path to the ScanNet CSV file on your side.

Pixie8888 commented 1 year ago

Thanks! By the way, I am running objectsdf++ on scannet scene 1 on one 3090Ti, but the training is very slow, probably 72h to finish 2000 epochs. Is it normal?

QianyiWu commented 1 year ago

I have updated the training code as in here. You don't need to run 2000 epochs on ObjectSDF++. In the original experiments, we set the total iteration as 200k and updated the nepochs.

Pixie8888 commented 1 year ago

oh Thanks! Do you have per scene result on ScanNet?

QianyiWu commented 1 year ago

You can refer to here.

Pixie8888 commented 1 year ago

Hi, I running objectsdf++ on ScanNet scene_1 using your updated codes, but the results is much worse than your given mesh results.

My results: Acc': 0.12937723174010277, 'Comp': 0.08671345835739624, 'Prec': 0.5227453147766669, 'Recal': 0.5302576311072802, 'F-score': 0.5264746758200511

Your provided mesh results: {'Acc': 0.041288216964188044, 'Comp': 0.05054920064427027, 'Prec': 0.7702384715344361, 'Recal': 0.672310470795879, 'F-score': 0.7179505307957033}

My experiment environment is pytorch2.0+cu11.7, python 3.9. Why does it have such big gap?

QianyiWu commented 1 year ago

Hi, this result looks strange. I checked my previous notes and found the reconstruction quality of scene_1 usually wouldn't meet such a low ACC value. But the environment would be one possible issue. The results I reported in my paper are under the environment of MonoSDF. When I prepare the open-source version, I add some additional supports such as vmap here and use torch2.0 for this code. I will try to check whether the environment matters that much.

joejep commented 4 months ago

I am trying to run the code on a new dataset. How can the dataset be generated