Skip to content

Latest commit

 

History

History
63 lines (49 loc) · 2.64 KB

README.md

File metadata and controls

63 lines (49 loc) · 2.64 KB

Data & Pre-trained Models

Please download the following datasets and pretrained models, and put them into the specified directory.

Preparing datasets

  • PASCAL-5i
  • COCO-20i

Dataset Structure

Final directory structure (only display used directories and files):

./data
├── COCO
│   ├── annotations
│   ├── train2014
│   ├── train2014_labels
│   ├── val2014
│   ├── val2014_labels
│   └── weights
├── VOCdevkit
│   └── VOC2012
│       ├── SegmentationClassAug
│       ├── JPEGImages
│       └── weights
└── README.md

PASCAL-5i

  • Download Training/Validation data (2G, tarball), and extract VOCtrainval_11-May-2012.tar to ./data/
  • Download SegmentationClassAug (34M, tarball, GoogleDrive or BaiduDrive Code: FPTr), and extract to ./data/VOCdevkit/VOC2012/. This is an extended annotation set from SBD.
  • Precomputed cross-entropy weights (only used for training)

    • Option 1: Download from BaiduDrive Code: FPTr, and extract pascal_weights.tar to ./data/VOCdevkit/VOC2012/. Rename the directory name to weights.

    • Option 2: Generate from datasets:

      # Dry run to ensure the output path are correct.
cuda 0 python tools.py precompute_loss_weights with dataset=PASCAL dry_run=True
# Then generate and save to disk.
cuda 0 python tools.py precompute_loss_weights with dataset=PASCAL

COCO-20i

  • Create directory ./data/COCO

  • Download 2014 Training images (13GB, zip), 2014 Val images (6GB, zip), 2014 Train/Val annotations (241M, zip), and extract them to ./data/COCO/

  • Generate offline labels

     python tools.py gen_coco_labels with sets=train2014
 python tools.py gen_coco_labels with sets=val2014

  • Precompute cross-entropy weights (only used for training)

    • Option 1: Download from BaiduDrive Code: FPTr, and extract coco_weights.tar to ./data/COCO/. Rename the directory name to weights.

    • Option 2: Generate from datasets:

      cuda 0 python tools.py precompute_loss_weights with dataset=COCO save_byte=True