data.md

Models & Data

Make sure you have the pre-trained detector-weight files and data prepared.

• Model: Pretrained detection model weights.
• Data: Dataset of images(required) & annotations(optional, for evaluation), and the corresponding detection labels (optional, generated by the victim models for evaluation).

Data

# Download image data and prepare correspond labels in directory data/.
# A demo file tree:
├── data
    ├── INRIAPerson
        ├── Train
        ├── Test
            ├── pos # data
            ├── labels # labels
                ├── faster_rcnn-rescale-labels
                ├── ground-truth-rescale-labels
                ├── ...

Download You can generate detection labels of your custom models (parsing COCO labels is also supported), see more details in README. see more details in README. Also, we provide to download our experimental label data and patches from GoogleDrive | BaiduCloud.

Data	Generated Labels	Source
CCTVPerson	GoogleDrive	Human Detection
COCOperson	GoogleDrive	HomePage
INRIAPerson	GoogleDrive	Paper

Models

You can download model weights by running:

# Run in the root proj dir.
# Download models.
bash ./detlib/weights/download.sh
# To create links to the corresponding detector modules.
bash ./detlib/weights/setup.sh

You can also download the supported models fom the aforementioned links. Your file tree of the downloaded weights should like this:

# The weight file tree should like this.
└── detlib
    ├── base.py
    ├── HHDet
    ├── torchDet
    └── weights
        ├── setup.sh
        ├── yolov2.weights
        ├── yolov3-tiny.weights
        ├── ...