noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation

MIT License

527 stars 58 forks source link

retrain #118

Closed haifengwu205 closed 6 months ago

haifengwu205 commented 6 months ago

Hi, I retrain your coder, however, some of the metrics dropped significantly. The following are the results of my retraining. sq_rel and rmse dropped significantly. I didn't modify any parameters, why is this happening?

abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |

& 0.107 & 0.805 & 4.591 & 0.183 & 0.889 & 0.962 & 0.983 \

noahzn commented 6 months ago

Hi, how did you evaluate the training? Did you only evaluate on the last epoch? There are some discussions regarding the reproducibility. https://github.com/noahzn/Lite-Mono/issues/7 https://github.com/noahzn/Lite-Mono/issues/4

noahzn commented 6 months ago

I'm closing this issue now.