Skip to content

Latest commit

 

History

History
82 lines (63 loc) · 2.79 KB

dataset.md

File metadata and controls

82 lines (63 loc) · 2.79 KB
Raw

Dataset

The vswarm drone detection dataset comprises 11822 images (9891 for training, 1931 for validation) recorded in five different environments around the EPFL campus. The images are taken from a static camera and automatically labeled using background subtraction (modulo some minor data cleaning).

The following camera hardware was used to acquire the images:

The tree structure of the dataset is as follows:

vswarm_drone_detection_dataset
├── <environment-name>_<flight-#>
│   ├── images
│   │   ├── frame_<seq-#>_<unix-timestamp>.png   ---> Image file
│   │   ├── frame_<seq-#>_<unix-timetsamp>.json  ---> Label file (LabelMe format)
│   │   ├── ...
│   ├── labels
│   │   ├── frame_<seq-#>_<unix-timestamp>.txt   ---> Label file (Darknet format)
│   │   ├── ...
├── ...
├── README.md  ---> This README
├── train.txt  ---> Paths of images used for training (9891 examples)
├── valid.txt  ---> Paths of images used for validation (1931 examples)

We provide labels in two different formats.

Darknet format

The Darknet format is .txt-based and mainly useful for training/testing the detector.

Each .txt file has the following format (one detection per line):

<class-id> <center-x> <center-x> <width> <height>

where:

  • <class-id> is the zero-indexed ID of the class (only one in our case: jetson_lequad)
  • <center-x/y> is the normalized (range 0-1) center of the bounding box in x/y-direction
  • <width/height> is the normalized (range 0-1) width/height of the bounding box

Example:

0 0.436806 0.699074 0.123611 0.150000

LabelMe format

The LabelMe format is .json-based and mainly useful for reviewing the auto-generated labels and/or manual labeling.

Each .json file has the following format:

{
    "fillColor": [255, 0, 0],
    "flags": {},
    "imageData": null,
    "imageHeight": 540,
    "imagePath": "frame_075_1599398126643726885.png",
    "imageWidth": 720,
    "lineColor": [0, 255, 0, 128],
    "shapes": [
        {
            "fill_color": null,
            "flags": {},
            "label": "jetson_lequad",
            "line_color": null,
            "points": [[270, 337], [359, 418]],
            "shape_type": "rectangle"
        }
    ],
    "version": "3.16.2"
}

where "points" specifies the top left and bottom right extreme point of the bounding box and "label" the class ID.