search

lhoyer / improving_segmentation_with_selfsupervised_depth

[CVPR21] Implementation of our work "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation"
247 stars 30 forks source link

Clarification regarding the semi supervised segmentation task #7

Closed testingshanu closed 3 years ago

testingshanu commented 3 years ago

  1. Which experiment id is suitable to train the semantic segmentation network using all the labels from cityscape ? (while using the monodepth pretrained model)

  2. Is "leftImg8bit_sequence_small" data required while running only the semi supervised segmentation experiments ?

lhoyer commented 3 years ago
  1. You can achieve that by enabling 2975 samples and disabling 372 samples in the subsets function in experiments.py. Further, you should comment our "(f'sel_{pres_method}_pad_transfer_dcompgt{dc_m}{dc_ft}', True, "depthcomp", False, True, True)" in experiments.py. After that, you can run python run_experiments.py --machine ws --exp 212. If you run into memory issues, please consider issue #4.
  2. Yes, the sequences are also required for the fully supervised case due to the MTL with self-supervised depth.
testingshanu commented 3 years ago

Thank you for the reply.

What does MTL indicate ?

lhoyer commented 3 years ago

MTL=Multi-Task Learning