MonoDepthSegReprod

Final project for CS 7643 Deep Learning focusing on reproducing the results detailed in "Self-Supervised Monocular Scene Decomposition and Depth Estimation" - S. Safadoust, F. Guney (2021)

Dependencies (Current)

• torchvision-0.11.1
• pytorch-1.10.0
• numpy-1.21.2

Training and Validation

Command for Training:

• python train.py --png --data_path ../kitti_data --load_weights_folder /home/luis/backup_model_run/mdp_run5/networks/weights_5/

Command for Evaluation:

• python depth_masking_validation.py --png --data_path ../kitti_data --load_weights_folder /home/luis/backup_model_run/mdp_run6/networks/weights_4/ --num_val_batches=2

Command for Viewing Logs:

• tensorboard --logdir=<dir_of_log>

References

This codebase reproduces the MonoDepthSeg architecture described in the following work:

Self-Supervised Monocular Scene Decomposition and Depth Estimation

Sadra Safadoust, Fatma Güney

3DV 2021 (arXiv pdf)

No public code implementations for this work currently exist.

---

This codebase also uses model architectures and utility functions from the following work per their non-commercial use license:

Digging into Self-Supervised Monocular Depth Prediction

Clément Godard, Oisin Mac Aodha, Michael Firman and Gabriel J. Brostow

ICCV 2019 (arXiv pdf)

nianticlabs/monodepth2 (Github)

Copyright © Niantic, Inc. 2018. Patent Pending.