README.md

Evaluation on LayoutBench-COCO

Evaluation of generated images on LayoutBench-COCO, with pretrained YOLOv7 model.

Setup Env

pip install -r requirements.txt
pip install pycocotools

Evaluation on LayoutBench-COCO

We provide evaluation scripts for the following models:

• GLIGEN - https://huggingface.co/masterful/gligen-1-4-generation-text-box
• ReCo (COCO checkpoint) - https://huggingface.co/j-min/reco_sd14_coco
• ControlNet - https://huggingface.co/lllyasviel/control_v11p_sd15_seg
• IterInpaint (COCO checkpoint) - https://huggingface.co/j-min/iterinpaint_sd15inpaint_coco

Please see ../image_generation/README.md for image generation scripts on LayoutBench-COCO.

models=(
    'reco'
    'gligen'
    'controlnet_seg'
    'iterinpaint'
)

for model in "${models[@]}"

do
    echo $model
    python prepare_image_fnames.py \
        --model $model \
        --layout_input_path "../image_generation/data/layoutbench-coco/all_layouts.json"
        --image_dump_base_dir "../image_generation/images/${model}"
        --image_list_dump_dir "./image_fnames_dir/${model}_all_layouts.txt"
done

Run evaluation

export RESULTS_OUT_PATH='results/layoutbench_real.json'

layout_names=(
    'combinations_layouts_all'
    'count_layouts_all'
    'position_layouts_all'
    'size_layouts_all'
)

models=(
    'reco'
    'gligen'
    'controlnet_seg'
    'iterinpaint'
)

for model in "${models[@]}"
do
    echo $model
    for layout_name in "${layout_names[@]}"
    do
        export image_paths=./image_fnames_dir/${model}_${layout_name}.txt
        export anno_json=./data/layoutbench-coco/${layout_name}.json
        export exp_name=${model}_${layout_name}

        echo $layout_fname
        echo $image_paths

        python test.py --data data/layoutbench_real.yaml --img 640 --batch 32 --conf 0.001 --iou 0.65 --device 0 --weights yolov7.pt \
            --name $exp_name \
            --image_paths $image_paths \
            --anno_json $anno_json \
            --save-json \
            --save_result_json_path $RESULTS_OUT_PATH
    done
done