search

ethz-asl / background_foreground_segmentation

12 stars 2 forks source link

Feature/weakly supervised training #17

Closed renezurbruegg closed 3 years ago

renezurbruegg commented 3 years ago

This will probably be a long pull request, so you might want to grab a coffee before you start :coffee: :)

There are a few main changes here:

  1. Changed the dataloader to load depth + filtered images: I created a separate pull request for this as i assume that some changes are needed: See #16

  2. Experiments.py I created a class Experiment.py which should provide all the information that is needed for a given experiment. Each new Experiment (e.g. SemSegWithDepth / SemSegWithPseudoLabel) can then overwrite specific functions (e.g. get loss for PseudoLabels). All experiments can request specific parameters using the argparser interface and the _addArguments() function.

  3. Losses I created new losses based on the papers i showed during the weekly meetings. Additionally i also created an Ignorant Wrapper to ignore NaN values inside the loss ground truth

  4. Models I added a DeeplabV3Plus model with depth/semseg prediction as i was having issues changing the PSP model. I plan to implement a multi task PSP model later.

renezurbruegg commented 3 years ago

I might have messed up merging the main branch before... Ignore all commits 250f066 - b06d994 :D

hermannsblum commented 3 years ago

@renezurbruegg since we are open-sourcing now, is there anything in here that we don't already have?

renezurbruegg commented 3 years ago

@hermannsblum Without digging through all the new code, I think the Depth Estimation + Consistency loss might be missing. However it's not really part of the paper and there was some additional work done on it iirc.