fengshikun/yolt

You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery

YOLT

You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery

YOLT is an extension of the YOLO framework that can evaluate satellite images of arbitrary size, and runs at ~50 frames per second. Current applications include vechicle detection (cars, airplanes, boats), building detection, and airport detection.

The YOLT code alters a number of the files in src/*.c to allow further functionality. We also built a python wrapper around the C functions to improve flexibility. We utililize the default data format of YOLO, which places images and labels in different directories. For example:

/data/images/train1.tif
/data/labels/train1.txt

Each line of the train1.txt file has the format

<object-class> <x> <y> <width> <height>

Where x, y, width, and height are relative to the image's width and height. Labels can be created with LabelImg, and converted to the appropriate format with the /yolt/scripts/convert.py script.

For more information, see:

1. arXiv paper: You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery

2. Blog1: You Only Look Twice — Multi-Scale Object Detection in Satellite Imagery With Convolutional Neural Networks (Part I)

3. Blog2: You Only Look Twice (Part II) — Vehicle and Infrastructure Detection in Satellite Imagery

4. Blog3: Building Extraction with YOLT2 and SpaceNet Data

5. Blog4: Car Localization and Counting with Overhead Imagery, an Interactive Exploration

6. Blog5: The Satellite Utility Manifold; Object Detection Accuracy as a Function of Image Resolution

7. Blog6: Panchromatic to Multispectral: Object Detection Performance as a Function of Imaging Bands

---

Installation

1. Install nvidia-docker

2. Build docker file

    nvidia-docker build -t yolt yolt_container /path_to_yolt/docker
   

---

Execution

Commands should be executed within the docker file. To run the container (with name yolt_train0):

nvidia-docker run -it -v --name yolt_train0 yolt_container

HELP

cd /path_to_yolt/scripts
python yolt2.py --help

COMPILE (gpu machine)

cd /path_to_yolt/scripts
python yolt2.py \
    --mode compile

TRAIN (gpu_machine)

# e.g. train boats and planes
cd /path_to_yolt/scripts
python yolt2.py \
    --mode train \
    --outname 3class_boat_plane \
    --object_labels_str  boat,boat_harbor,airplane \
    --cfg_file ave_13x13.cfg  \
    --nbands 3 \
    --train_images_list_file boat_airplane_all.txt \
    --single_gpu_machine 0 \
    --keep_valid_slices False \
    --max_batches 60000 \
    --gpu 0

VALIDATE (gpu_machine)

# e.g. test on boats, cars, and airplanes
cd /path_to_yolt/scripts
python yolt2.py \
    --mode valid \
    --outname qgis_labels_all_boats_planes_cars_buffer \
    --object_labels_str airplane,airport,boat,boat_harbor,car \
    --cfg_file ave_26x26.cfg \
    --valid_weight_dir train_cowc_cars_qgis_boats_planes_cfg=ave_26x26_2017_11_28_23-11-36 \
    --weight_file ave_26x26_30000_tmp.weights \
    --valid_testims_dir qgis_validation/all \
    --keep_valid_slices False \
    --valid_make_pngs True \
    --valid_make_legend_and_title False \
    --edge_buffer_valid 1 \
    --valid_box_rescale_frac 1 \
    --plot_thresh_str 0.4 \
    --slice_sizes_str 416 \
    --slice_overlap 0.2 \
    --gpu 2

---

To Do

1. Include train/test example
2. Upload data preparation scripts
3. Describe multispectral data handling
4. Describle initial results with YOLOv3
5. Describe improve labeling methods

---

If you plan on using YOLT in your work, please consider citing YOLO and YOLT