DetectiveKite

DetectiveKite is a free perch mount camera trap images auto-detection system which is developed base on yolov5s, a light and fast object detection architecture. I believe DetectiveKite can help biologists in Taiwan to accelerate the heavy species labeling works.

Datasets

Transfer Learning

• NABirds Dataset

Augmentation

• Private data from Endemic Species Research Institute
• Landscape Pictures, Kaggle
  • Owner: Arnaud ROUGETET
• Landscape color and grayscale images, Kaggle
  • Owner: Black Mamba

Train and Validation

• Private data from Institute of Wildlife Conservation, National Pingtung University of Science and Technology
• Private data from Department of Biological Resources, National Chiayi University
• iNaturalist

Install

Environment

Python

yolov5 need python>=3.7.0, DetectiveKite is developed under 3.9.10

Others

Please refer to the "FFmpeg 安裝" and "安裝 CUDA、cuDNN" of SILIC 環境設定 for Window 10 or 11

git clone

git clone https://github.com/Chendada-8474/DetectiveKite.git
cd DetectiveKite

Packages

pip install -r requirements.txt

GPU Accelerated Computing

For GPU acceleration, please install the compatible torch package with your device. see INSTALL PYTORCH. Futhermore, installing CUDA Toolkit correctly is also necessary.

Open in Colab

Colab is an easy way to execute python on browser.

• setting friendly
• free GPU
• easy to share code

Documentation

Detection

detect.py

detect.py runs source of a folder. use --source or -so to set path of target folder. The results will be saved to runs/data/.

python detect.py -so path\      # directory

Argument Parameter

Trail camera normally fill infrared light at night or in bad light condition when animals are detected and take a black and white photo. Otherwise, it takes color photo in Natural light. Two models are used to detect color and infrared image saperatly. So, please note that some parameter have to be set saperatly.

argument	abbreviation	parameter	default	require
--source	-so	path		True
--batch-size	-bs	batch size, positive integer		True
--classes-color	-cc	int series	all classes	False
--classes-infrared	-ci	int series	all classes	False
--conf-thres-color	-ctc	float, range 0~1	0.25	False
--conf-thres-infrared	-cti	float, range 0~1	0.25	False
--vedio-interval	-vi	float (second)	1	False
--name	-na	results folder name	exp	False
--sp-name	-sn	sci, ch or jp	en	False

Batch Size

Please set this as big as you can.

python detect.py -so path\ -bs 24      # Batch Size 24

Classes

When deteting, you are able to use --classes-color and --classes-infrared to spefify classes. Information of classes is in ./model/exp_color/classes_color.csv and ./model/exp_infrared/classes_infrared.csv.

python detect.py -so path\   # all classes

color

python detect.py -so path\ -cc 0      # only specify class o
                               0 15   # specify class 0 and 15

infrared

python detect.py -so path\ -ci 0      # only specify class o
                               0 3    # specify class 0 and 3

or both

python detect.py -so path\ -cc 0 15 -ci 0 3

Confidence Threshold

Setting the confidence threshold when detecting using --conf-thres-color and conf-thres-infrared. The results which confidence is below than confidence threshold will not be saved.

python detect.py -so path\ -ctc 0.6 -cti 0.75

Video Interval

Detecting EACH frame of video just for knowing what kind of birds stanted on perch mount is not a efficient way. DetectiveKite detect video in interval to speed up the processing of detection. Set --video-interval to adjust the suitable detection interval base on your video length.

python detect.py -so path\ -vi 2     # detect video frame only in 2s interval

Saved Folder Name

After detection, results will be saved to ./runs/data/exp/. exp is the default folder name. You can change the name via --name.

python detect.py -so path\ -na project

Then, results will be saved to ./runs/data/project/

Species Name

The species name is saved in English common name (default). You can choose in scientific name, Chinese common name, Japanese common or Multiple.

name	lang code
scientific name	sn
Chinese common name	ch
Japanese common name	jp

non for English common name

python detect.py -so path\

add name

python detect.py -so path\ -sn ch           # add Chinese common name
　　　　　　　　　　　　　　　　　sci          # add scientific name
                               jp           # add Japanese name
                               ch sci jp    # Multiple is also legal

Draw Bounding Box and Label Name

If you want some thing cool like these:

Please use yolov5 detect to do it. see yolov5 package

or video

• Color Video detection Demo
• Infrared Video detection Demo

example:

yolov5 detect --source ./sample/img019.jpg --weights ./model/exp_color/best.pt    # color media

yolov5 detect --source ./sample/img001.jpg --weights ./model/exp_infrared/best.pt    # infrared media

Results

Results Format

Results are saved as .csv format. Columns: file_name, class, name, num_inds, confidence, media, model.

Information of columns:

column	data type	description
file_name	string	file name of image or video
class	integer	code of class
name	string	name of detected species
num_inds	integer	number of individuals detected
confidence	float	see this
datetime	datetime	file create datetime
media	string	image or video
model	string	detected by color model or infrared model
xmin	integer	minimum pixel of bounding box on x axis
ymin	integer	minimum pixel of bounding box on y axis
xmax	integer	maximum pixel of bounding box on x axis
ymax	integer	maximum pixel of bounding box on y axis

NOTE: If detected media is video, the confidence is the average of every detected frame. In image, if two or more individuals detected, the xmin and ymin are the smallest values of xmin and ymin. The xmax and ymax are the biggest values of xmax and ymax. In video, the xmin and ymin are the smallest values of xmin and ymin of all frames. The xmax and ymax are the biggest values of xmax and ymax of all frames.

Review

review.py is a GUI interface results reviewer.

python review.py

Hotkey

Key	Description type
A	previous media
D	next media
L	add species
C	confirm changes
1 ~ 9	focus to the species input box of row 1 ~ 9
Ctrl + 1 ~ 9	delete the species input box of row 1 ~ 9
Ctrl+S	save csv file

NOTE: If some video formats are not supported, try installing K-lite.

Donation

Althought, this is not a big project, I maintain this project without any financial or funds support. If you think this is great and want to encourage me, donation would be so helpful!

In addition, if you have tons of perch mount images and videos are waited to be labeled, we could cooperate! contact me!

綠界科技

https://p.ecpay.com.tw/0696F33

PayPal

https://paypal.me/tachihchen