Available trained models

cd "<GENERAL_EXP_OUTPUT_DIR>" # TODO replace!

arr=(
    ret_distilbert_dataPSA2ftPSH2
    modret_distilbert_dataPFAftPFH
    gen_distilbert_dataPSA2ftPSH2
    b_gen_distilbert_dataPFAftPFH
    capgen_CAtransfPSA2H2_dataPSA2ftPSH2
    modgen_CAtransfPFAHPP_dataPFAftPFH
)

for a in "${arr[@]}"; do
    echo "Download and extract $a"
    wget "https://download.europe.naverlabs.com/ComputerVision/PoseFix/${a}.zip"
    unzip "${a}.zip"
    rm "${a}.zip"
done

• Text-to-pose retrieval: providing an improved model, pretrained on new automatic captions, and with a symmetric constrastive loss (vs. uni-directional contrastive loss in the paper)
• Instruction-to-pair retrieval: providing an improved model trained with symmetric contrastive loss (vs. uni-directional contrastive loss in the paper).
• Pose editing: the provided model uses a transformer-based text encoder (frozen DistilBert + learned transformer), for consistency with the other provided models (vs. GloVe+biGRU configuration used to report results in the paper). Note: this model was finetuned using the best setting as per Table 4: with L/R flip and paraphrases. The FID value may also change as evaluation is carried out with an improved version of the text-to-pose retrieval model.
• Text generation models: evaluated with improved retrieval models; also note that, despite an average over 10 repetitions, R-precision metrics come with a great variability due to the randomized selection of the pool of samples to compare against.