Darknet Video

• Extract python APIs from darknet to maintain python wrapper easier.

Object Detection and Tracking

Robot Interaction

Use with python-OP3

Gesture Recognition

Head Tracking

Feature

• Video object tracking using SiamMask
• Global NMS for all object (not each class)
• Object size threshold and class filtering
• MJPEG streaming for YOLO+SiamMask result on port 8091
• Auto labeling powered by YOLO+SiamMask.

Requirements

• Compiled OpenCV with python wrapper and TrueType contrib module

You can download below files in Releases, and unzip to project root:

• Compiled darknet yolo_cpp_dll.dll
• CUDA Toolkit 10.1 + cuDNN 7.6.5
• the darknet weight you want to use (ex: yolov4.weight)

Development Installation

pip3 install --user -e .

Usage

from darknet_python.run_detector import ThreadingDetector
# First argument same as cv2.captureVideo() argument, here we use web camera.
# If not pass config_path, loading config of basename of weights by default.
# By default, it load meta file of coco.data.
ThreadingDetector(0, weights_path='yolov3.weights')

You can find additional examples on test/test_darknet.py.

Pyinstaller

You can use pyinstaller to generate standalone executable

pyinstaller -F --add-binary "../lib;lib" --add-data "../darknet_video/cfg;darknet_video/cfg" --add-data "../darknet_video/data;darknet_video/data" --add-data "../weights;weights" .\test_darknet.py

Labeling assisted by YOLO-SiamMask

Automating labeling use native detection ability of YOLO and SiamMask.

To-do list

• click to select yolo box to candidate label
• frame by frame human check
• manual bbox selecting
• adjust yolo predict box