Skip to content

Latest commit

 

History

History
55 lines (50 loc) · 2.14 KB

DATA.md

File metadata and controls

55 lines (50 loc) · 2.14 KB

Data Preparation

Our model was trained using three datasets: Objects365v1, V3Det, and OpenImages. We conducted tests on the LVIS dataset in a zero-shot manner. Please organize the datasets as follows.

Pretrained Weights

SAM Pretrain Weights (ViT-base)

mkdir -p sam_checkpoints
cd sam_checkpoints
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
cd ..

Data

Training

  1. Datasets preparation Download the datasets from their respective official websites. Ensure that you have objects365, V3Det and OpenImages V6. Organize the downloaded datasets as follows:
${ROOT}
    -- datasets
        --objects365
        --v3det
        --openimages
  1. Mask Token Preparation As the SAM (Segment Anything Model) has been set to a frozen state, we've optimized our resource usage by pre-extracting the image mask tokens. This step significantly reduces memory consumption during model training and inference. We have made these pre-extracted mask tokens available for easy access: Download Masks Tokens from One Drive We anticipate the data to be organized as follows:
${ROOT}
    -- datasets
        -- datasets_mask_tokens_vit_b
            --objects365
            --v3det
            --openimages

Evaluation

For model evaluation, download the LVIS dataset from COCO, LVIS Dataset and place it in the datasets folder at the project root:

${ROOT}
    -- datasets
        --coco
        --lvis

After downloading the LVIS dataset, also obtain the bounding box results from GLIP by downloading the provided JSON file:

Once downloaded, place the JSON file in the glip_results directory within datasets:

${ROOT}
    -- datasets
        --glip_results
            nms_results_glip_tiny_model_o365_goldg_cc_sbu_lvis_val.json