/Layout-based-sTDE

Layout-based Causal Inference for Object Navigation (CVPR 2023)

Primary LanguagePythonMIT LicenseMIT

Layout-based Causal Inference for Object Navigation

Setup

  • Clone the repository and move into the top-level directory cd Layout-based-sTDE
  • Create conda environment. conda env create -f environment.yml
  • Activate the environment. conda activate ng
  • We provide pre-trained model of ORG+L-sTDE. Please download and put it in the ./trained_models folder.
  • Our settings of dataset follows previous works, please refer to HOZ for AI2THOR and SemExp for Gibson.

Training and Evaluation

Train our Layout-based model

python main.py \
      --title layoutmodel \
      --model LayoutModel \
      --workers 12 \
      --gpu-ids 0

Evaluate our model with sTDE (our Layout-based sTDE model)

python full_eval.py \
        --title layoutmodel \
        --model LayoutModel \
        --results-json layoutmodel_sTDE.json \
        --gpu-ids 0 \
        --TDE_self True \
        --TDE_threshold 0.5 \
        --TDE_mode zero