Failed to reproduce paper results

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020

MIT License

321 stars 55 forks source link

Failed to reproduce paper results #20

Closed BrennoCaldato closed 3 years ago

BrennoCaldato commented 3 years ago

Hi @zzangjinsun. Thanks for sharing the code.

I tried to reproduce the paper results with Kitti and NYU datasets, using the pre-trained weights you provided.

I followed the Readme instructions but the output depth images are not accurate. Here are the quantitative results I'm getting for the NYU dataset:

Metric | RMSE: 0.0922 MAE: 0.0348 iRMSE: 0.0139 iMAE: 0.0050 REL: 0.0117 D^1: 0.9955 D^2: 0.9993 D^3: 0.9999

Did you face this issue before?

zzangjinsun commented 3 years ago

Hi @BrennoCaldato , I think you got the correct results, but D^1, D^2, and D^3 are not in percent so you can multiply 100 to compare with the results in the paper.