basegun-ml v2.0.0

Changelog:

Add the OCR module with alarm model detection

Add exception management when a card or a firearm is absent in the measure module

Change the classification model from nano to small

Deep Learning Model version

Classification model

The classification model is based on YOLOV8.

model size : Small

dataset used : basegun V1

data augmentation : flip LR : 0.2, flip UD: 0.2, rotate +-180

hyperparameters : adamW, batch size 16

training date : January 15th 2024

Keypoint detection model

The keypoint detection model is based on YOLOV8.

model size : nano

dataset used : keypoints V1

data augmentation : None to avoid creation of "false" images

hyperparameters : Loss function : euclidean distance instead of OKS, Optimizer AdamW

training date : October 27th 2023

Oriented Bounding Box card detection model

The Oriented Bounding Box card detection model is based on YOLOV5.

model size : nano

dataset used : Card detection V1

data augmentation : No flip because of bounding box order, HSV augmentation full, rotate +-180

hyperparameters : batch size 2, optimizer: Adam

training date : December 4th 2023

OCR models

The OCR model is based on paddleOCR

recognition and detection models: ch_PP-OCRv4

classification mode ch_ppocr_mobile_v2.0_cls_infer

IQA model

the IQA model is based on CNNIQA

basegun-ml v1.0.1

First version of the basegun-ml package

Deep Learning Model version

Classification model

The classification model is based on YOLOV8.

model size : nano

dataset used : basegun V1

data augmentation : flip LR : 0.2, flip UD: 0.2, rotate +-180

hyperparameters : adamW, batch size 16

training date : January 15th 2024

Keypoint detection model

The keypoint detection model is based on YOLOV8.

model size : nano

dataset used : keypoints V1

data augmentation : None to avoid creation of "false" images

hyperparameters : Loss function : euclidean distance instead of OKS, Optimizer AdamW

training date : October 27th 2023

Oriented Bounding Box card detection model

The Oriented Bounding Box card detection model is based on YOLOV5.

model size : nano

dataset used : Card detection V1

data augmentation : No flip because of bounding box order, HSV augmentation full, rotate +-180

hyperparameters : batch size 2, optimizer: Adam

training date : December 4th 2023

Dataset v0

• add functions for testing images in post_training.py
• regorganize models folder
• write more details about training params
• keep lr in checkpoints
• do not shuffle val dataset