Skip to content

Latest commit

 

History

History
302 lines (213 loc) · 7.43 KB

data.md

File metadata and controls

302 lines (213 loc) · 7.43 KB

Data Preparation

It is recommended to symlink the dataset root to data/ under the project root directory. If your folder structure differs, you will need to modify the data_root path in the corresponding config files.

Data preparation scripts are located in tools/data_prepare/. Remember to UNZIP it.

Download Datasets

We recommend downloading datasets from OpenDataLab for faster download speeds. Alternatively, refer to tools/data_prepare/<dataset_name>/README.md for official download links.

Supported Datasets

Dataset Classes Image Size Annotation Format Config
DIOR 20 800x800 VOC-style txt configs/_base_/datasets/dior.py
DOTA v1.0 15 1024x1024 (after split) DOTA txt configs/_base_/datasets/dota.py
DOTA v1.5 16 1024x1024 (after split) DOTA txt configs/_base_/datasets/dotav15.py
STAR 48 1024x1024 (after split) DOTA txt configs/_base_/datasets/star.py
FAIR1M - 1024x1024 (after split) DOTA txt configs/_base_/datasets/fair.py
HRSC 1 - VOC-style xml configs/_base_/datasets/hrsc.py
SSDD 1 - DOTA txt configs/_base_/datasets/ssdd.py
HRSID 1 - COCO json configs/_base_/datasets/hrsid.py
SRSDD 6 - DOTA txt configs/_base_/datasets/srsdd.py
RSDD 2 - VOC-style xml configs/_base_/datasets/rsdd.py

DIOR

Directory Structure

Point2RBox-v3
├── data
│   ├── dior
│   │   ├── JPEGImages-trainval/      # Train + val images
│   │   ├── JPEGImages-test/          # Test images
│   │   ├── Annotations/
│   │   │   ├── Oriented Bounding Boxes/   # Rotated box annotations
│   │   │   └── Horizontal Bounding Boxes/ # Horizontal box annotations
│   │   └── ImageSets/
│   │       └── Main/
│   │           ├── train.txt
│   │           ├── val.txt
│   │           └── test.txt

Config Path

Set data_root to data/dior/.

DOTA v1.0

Raw Directory Structure

Point2RBox-v3
├── data
│   ├── DOTA
│   │   ├── train/
│   │   ├── val/
│   │   └── test/

Image Splitting

DOTA raw images can be very large (up to thousands of pixels). They must be split into patches before training.

Single-scale split (1024x1024 patches with 200-pixel overlap):

# Split trainval
python tools/data_prepare/dota/split/img_split.py --base-json \
  tools/data_prepare/dota/split/split_configs/ss_trainval.json

# Split test
python tools/data_prepare/dota/split/img_split.py --base-json \
  tools/data_prepare/dota/split/split_configs/ss_test.json

Multi-scale split (if multi-scale training is needed):

python tools/data_prepare/dota/split/img_split.py --base-json \
  tools/data_prepare/dota/split/split_configs/ms_trainval.json

python tools/data_prepare/dota/split/img_split.py --base-json \
  tools/data_prepare/dota/split/split_configs/ms_test.json

Note: Before splitting, update the img_dirs and ann_dirs fields in the JSON config files to your actual raw data paths.

Directory Structure After Splitting

Point2RBox-v3
├── data
│   ├── split_ss_dota/
│   │   ├── trainval/
│   │   │   ├── images/       # Split train+val images
│   │   │   └── annfiles/     # Annotation files (one txt per image)
│   │   └── test/
│   │       ├── images/       # Split test images
│   │       └── annfiles/

Generate COCO-Format Annotations (Optional)

If you need COCO-format annotations:

python tools/data_prepare/dota/dota2coco.py \
  data/split_ss_dota/trainval/ \
  data/split_ss_dota/trainval.json

python tools/data_prepare/dota/dota2coco.py \
  data/split_ss_dota/test/ \
  data/split_ss_dota/test.json

Config Path

Set data_root to data/split_ss_dota/.

DOTA v1.5

Similar to DOTA v1.0, but with an additional container-crane class (16 classes total). The download and splitting procedures are the same; just use DOTA v1.5 data.

Config Path

Set data_root to data/split_ss_dotav15/.

STAR

Image Splitting

STAR also requires image splitting. The procedure is similar to DOTA. Refer to the DOTA splitting scripts for guidance.

Directory Structure After Splitting

Point2RBox-v3
├── data
│   ├── split_ss_star/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── annfiles/
│   │   ├── val/
│   │   │   ├── images/
│   │   │   └── annfiles/
│   │   └── test/
│   │       └── images/

Config Path

Set data_root to data/split_ss_star/.

FAIR1M

Image Splitting

FAIR1M requires image splitting:

# Split trainval
python tools/data_prepare/fair/split/img_split.py --base-json \
  tools/data_prepare/fair/split/split_configs/ss_trainval.json

# Split test
python tools/data_prepare/fair/split/img_split.py --base-json \
  tools/data_prepare/fair/split/split_configs/ss_test.json

Config Path

Set data_root to data/split_ss_fair/.

HRSC

Directory Structure

Point2RBox-v3
├── data
│   ├── hrsc/
│   │   ├── FullDataSet/
│   │   │   ├── AllImages/        # All images
│   │   │   ├── Annotations/      # XML annotations
│   │   │   ├── LandMask/
│   │   │   └── Segmentations/
│   │   └── ImageSets/            # Data split files

Config Path

Set data_root to data/hrsc/.

SSDD

Directory Structure

Point2RBox-v3
├── data
│   ├── ssdd/
│   │   ├── train/
│   │   └── test/
│   │       ├── all/              # All test images
│   │       ├── inshore/          # Inshore scenes
│   │       └── offshore/         # Offshore scenes

Config Path

Set data_root to data/ssdd/.

HRSID

Directory Structure

Point2RBox-v3
├── data
│   ├── HRSID_JPG/
│   │   ├── JPEGImages/           # All images
│   │   └── annotations/          # COCO-format JSON annotations

Config Path

Set data_root to data/HRSID_JPG/.

SRSDD

Directory Structure

Point2RBox-v3
├── data
│   ├── srsdd/
│   │   ├── train/
│   │   └── test/

Config Path

Set data_root to data/srsdd/.

RSDD

Directory Structure

Point2RBox-v3
├── data
│   ├── rsdd/
│   │   ├── Annotations/          # XML annotations
│   │   ├── ImageSets/            # Data split files
│   │   ├── JPEGImages/           # Training images
│   │   └── JPEGValidation/       # Validation images

Config Path

Set data_root to data/rsdd/.

Using Symlinks

If your datasets are stored elsewhere (e.g., shared storage), symlinks are recommended:

# Create data directory under project root
mkdir -p data

# Create symlinks (modify paths accordingly)
ln -s /path/to/your/dior           data/dior
ln -s /path/to/your/split_ss_dota  data/split_ss_dota
ln -s /path/to/your/split_ss_star  data/split_ss_star
ln -s /path/to/your/hrsc           data/hrsc
# ... similarly for other datasets