AutomatedAnnotations

Create annotations automatically from videos and save them in YOLO object detection format

Description

This project has two scripts "automated_annotation.py" and "automated_multi_class_annotation.py"

Single Class annotation (automated_annotation.py)

This python script works only when:

1. There is only one class

2. There is only one object of the class in the video

Multiple Class annotation (automated_multi_class_annotation.py)

User can interact with the script using command prompt

The output format of annotations is in YOLO object detection Format

class_id center_x center_y width height

Each bounding box is represented in the above format. All coordinates are in normalized representation (divided by width and height of the image)

Each image contains each bounding box in the above format in a new line

The project is created using the correlation tracker of dlib

I used opencv to both interact with and display the bounding boxes of the object in the video

Requirements

1. opencv
2. dlib

Usage

Usage Video

The demo video used contains an Elephant walking around.

The original video can be found at https://www.videezy.com/abstract/41648-elephant-walking-in-cage-at-zoo-4k

single class annotation

multiple class annotation

Command Prompt Commands

python3 automated_annotation.py -i input_video.mp4 -s annotations_newer -n 10 -o 5

The arguments are:

-i -> input video name

-s -> the annotations folder where two folders called images/ and annotations/ are created to store the annotations

-n -> every n-th frame of the video is saved in the annotations folder

-o -> the starting of the sequence number so that mixing the datasets become easier

-w -> width of the window (default=800px)

(only for multiple classes)

--small_object -> threshold for small object, the lower the value, the smaller the object

--frame_delay -> delay between two consecutive frames in ms (default = 1)

--start_frame -> starting frame in the video (default = 1)

--skip_frames -> number of frames to skip (default = 1)

-c -> classes that are separated by ','

Example:

python automated_multi_class_annotation.py -i Elephant.mp4 -c "elephant,tree" -s annotations_test -n 10

Steps:

1. The first frame is paused, simply click on the image displayed in the window titled "Automated Labelling", click and drag on the object to annotate.

2. (for multiple class only, skip if you want to try single class object annotation) Enter the name of the class in command prompt if you are using the multi class annotation script, and repeat the drag and annotate process.

3. Now click on the key 'p' on keyboard, the tracked object is displayed in green rectangle

4. When the tracking becomes inaccurate, click on the key 'p' on the keyboard again

5. Double-click right mouse button to delete the bounding box

6. Draw the updated bounding box

7. Select the class names in popup box in case of multiple object tracking

8. Press the key 'p' on keyboard to continue tracking

9. Press 'q' to stop the script execution.