Skip to content
/ yolt Public
forked from avanetten/yolt

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

License

Notifications You must be signed in to change notification settings

CosmiQ/yolt

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

44 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

YOLT

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

Alt text


As of 24 October 2018 YOLT has been superceded by SIMRDWN


YOLT is an extension of the YOLO v2 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

The following has been tested on Ubuntu 16.04.2

  1. Install nvidia-docker

  2. Build docker file

     nvidia-docker build -t yolt yolt_docker_name /path_to_yolt/docker
    
  3. Launch the docker container

     nvidia-docker run -it -v /raid:/raid yolt_docker_name
     # '/raid' is the root directory of your machine, which will
     # be shared with the docker container
    
  4. Run Makefile

     cd /path_to_yolt/
     make clean
     make
    

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_docker_name

HELP

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

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_standard.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_standard.cfg \
    --valid_weight_dir train_cowc_cars_qgis_boats_planes_cfg=ave_26x26_2017_11_28_23-11-36 \
    --weight_file ave_standard_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 YOLO v3
  5. Describe improve labeling methods

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

About

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

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C 79.2%
  • Python 7.8%
  • C++ 6.4%
  • Cuda 6.3%
  • Makefile 0.2%
  • Dockerfile 0.1%