Google Drones Dataset

Drone Detection Dataset from Google Images

1. Download Raw Images

See google-images-download for details.

googleimagesdownload -cf cfg.json

It is recommended that the .png files be converted to .jpg files for storage:

mogrify -format jpg -transparent-color white -background white -alpha background -flatten *.png
rm *.png

2. Filter Duplicates by Perceptual Hash

python unique.py

3. Bootstrapping : Create Initial Annotations

Option 1 : DarkNet

See drone-net for obtaining initial weights and parts of the dataset.

(Note that darknet must be compiled with an alternate version of detector.c to produce annotations.)

mkdir -p /tmp/det
./darknet detector test cfg/drone.data cfg/yolo-drone.cfg weights/yolo-drone.weights "/media/ssd/datasets/drones/all/"

Option 2 : From Pre-Trained Network

Alternatively, you may choose to produce the annotations from my pre-trained network.

After downloading the network and extracting to tmp/model, An example execution configuration might look like the following:

python box_ann.py -h
python box_ann.py --root=/tmp --model=model --use_gpu --noviz --img_dir=/media/ssd/datasets/drones/all --out_dir=/tmp/det --batch_size=16

The resultant annotations will be stored in /tmp/det as .txt files.

The format of the output .txt file is as follows:

filename    # full image path
cx cy w h p # text header, field names
cx cy w h p # one row per box, in relative coordinates. p=confidence in [0-1] interval

4. Manually Inspect and correct the annotations

First, record the responses:

python fix_ann.py [pos:=true]

The resultant responses will be saved in four different files:

ann_neg_idx.npy
ann_neg_lbl.npy
ann_pos_idx.npy
ann_pos_lbl.npy

Then apply the responses:

python apply_responses.py

This should store the results in dataset/

5. Convert To TFRecords

python tfrec.py

6. Training

python train.py --logtostderr --train_dir=training_demo/training/ --pipeline_config_path=pipeline.config

7. Export model

bash export_model.sh

8. Testing

python object_detection_tf.py