Projekt KNN

Téma: Odhad absolútnej hĺbky z obrazu (na prírodných dátach)

Autoři: Adam Ferencz, Tomáš Mojžiš, Petr Mičulek

Useful links

• Milanote tabule
• cphoto depth paper
• GeoPose3k
• megadepth paper

Project structure

Some parts of this project are taken from other repositories:

• semseg/ - this directory is taken from https://github.com/CSAILVision/semantic-segmentation-pytorch and it's not used in our implementation, we only used it at the beginning
• geopose/model/houglass_ugly.py - contains the original model from https://github.com/zhengqili/MegaDepth, we rewrote it because it was pretty ugly
• geopose/model/hourglass.py - contains the same model, but implemented in pytorch from scratch, inspired by https://github.com/dfan/single-image-surface-normal-estimation
• geopose/model/hourglass_*.py - contains derived models from base HG with added FOV
• geopose/train.py - Working with DistributedDataParallel https://gist.github.com/sgraaf/5b0caa3a320f28c27c12b5efeb35aa4c

Other files are scripts implemented by us. Main scripts are:

• geopose/train.py - training script, containing also evaluating on test dataset
• geopose/dataset.py - script containing DataLoader class
• geopose/losses.py - script containing implementation of loss functions
• and other minor utility scripts

Notes

Run from root repo directory (KNN-Projekt).

Loading JPEG uses TurboJPEG library that must be installed through a package manager.

sudo apt install libturbojpeg

Python package dependencies can be installed from requirements.txt.

pip install -r requirements.txt

How to launch

• install requirements.txt with pip
• for training on datasets/geoPose3K_final_publish run python geopose/train.py
  • set the parameters inside the script
• for plotting run python plot_predictions.py - set the weights_path variable
  • it could not be passed as parameter because of megadepth model builder uses argparse and it interferences