Tomster recognition with Darknet

1. Collecting the data

• images/photos
• movie

ffmpeg -i INPUT -filter:v fps=FRAMES/SECONDS %02d.jpg

2. Scaling

500px x 500px MAX

convert -resize 25% *.jpg

3. Labeling

https://github.com/puzzledqs/BBox-Label-Tool

4. Convert to VOC format

• script here

5. Spliting dataset

10% testing data

• script here

6. Create input file

• script here

7. Configuration

• tomster.names(labels/classes)

tomster

• tomster.data(sources)

classes = 1
train = /abs/path/to/train.txt
valid = /abs/path/to/test.txt
names = /abs/path/to/*.names
backup = backup/

• tomster.cfg(hyperparams)

soon

8. Training

Minimize AVG loss > 0.2

darknet detector train path/to/*.data path/to/*.cfg path/to/*.weights >> training.log

9. Testing

darknet detector demo path/to/*.data path/to/*.cfg path/to/*.weights

Change the detection threshold(optional):

thresh 0.X