yeyingdege / refcoco_data_tool Goto Github PK

To install requirements:

pip install -r requirements.txt

The data/ folder should look like this:

data
├── refcoco
|   ├── anns
|   |   ├── refcoco
|   |   ├── refcoco+
|   |   ├── refcocog
│   ├── images
|   │   ├── train2014  # images from train 2014
│   ├── masks
|   |   ├── refcoco
|   |   ├── ...

1. Download MSCOCO and link the train2014 folder to data/refcoco/images/train2014.

2. Download refcoco/masks and link it to data/refcoco/masks

3. Download refcoco/anns and put it under data/refcoco/

1. Run the following script to process RefCOCO dataset. It generates new annotation files under data/refcoco/anns_spatial

python datasets/build_spatial_anno.py

2. New dataloader (under development): datasets/SpatialReasoningDataset.py

Run datasets/refer_segmentation.py at root directory to dive into the dataloader and visualize annotations.