YALTAi

You Actually Look Twice At it

This provides an adapter for Kraken to use YOLOv5 Object Detection routine.

This tool can be used for both segmenting and conversion of models.

Routine

Instal

pip install YALTAi

Training

Convert (and split optionally) your data

# Keeps .1 data in the validation set and convert all alto into alto
#  Keeps the segmonto information up to the regions
python -m yaltai.yaltai alto-to-yolo PATH/TO/ALTOorPAGE/*.xml my-dataset --shuffle .1 --segmonto region

And then train YOLO (note that I recommend using the repository and not the CLI) as the CLI provided with the library keeps for looking at the wrong place (it needs absolute path)

# Train your YOLOv5 data (YOLOv5 is installed with YALTAi)
yolov5 train --data "$PWD/my-dataset/config.yml" --batch-size 4 --img 640 --weights yolov5x.pt --epochs 50

Predicting

YALTAi has the same CLI interface as Kraken, so:

• You can use base BLLA model for line or provide yours with -m model.mlmodel
• Use a GPU (--device cuda:0) or a CPU (--device cpu)
• Apply on batch (*.jpg)

# Retrieve the best.pt after the training
# It should be in runs/train/exp[NUMBER]/weights/best.pt
# And then annotate your new data with the same CLI API as Kraken !
python -m yaltai.kraken_yaltai --device cuda:0 -I "*.jpg" --suffix ".xml" segment --yolo runs/train/exp5/weights/best.pt

Metrics

The metrics produced from various libraries never gives the same mAP or Precision. I tried

• object-detection-metrics==0.4
• mapCalc
• mean-average-precision which ended up being the chosen one (cleanest in terms of how I can access info)

and of course I compared with YOLOv5 raw results. Nothing worked the same. And the library YOLOv5 derives its metrics from is uninstallable through pip.