adnan33/Glue_Tube_Length_Keypoint_Detection

This project attempts to predict glue tube length from images by keypoint detection.

Glue Tube Length Prediction Using KeyPoint Detection

This project attempts to predict glue tube length from keypoint detection of tube head and tail. As most of the images have two tubes present in the image, used model has eight(8) nodes in the logit layer. A simple data imputation was used to avert input error while training by inserting (0,0),(0,0) co-ordinates in images having single tube.

Image Augmentation

The following transformation techniques were used from Albumentations library

• Rescale to 512 for width and height with application probability of 1.0
• VerticalFlip with a application probability of 0.7
• HorizontalFlip with a application probability of 0.7
• HueSaturationValue with a application probability of 0.5
• RGBShift with application probability of 0.7
• RandomBrightnessContrast with application probability of 0.5
• Normalize & Convert ToTensor with application probability of 1.0

Architecture

A simple model with:

1. Three (3) Conv2D layers
2. ReLU Activation and MaxPooling in every layer
3. Two (2) fully-connected layers, and Dropout to prevent overfitting.

Hyperparameters

BatchSize, Epochs, Loss & Optimization Functions(using GPU)

• BatchSize : 8
• Epochs : 100 (can train longer for better performance)
• Loss : Mean squared error (MSE)
• Optimizer : Adam
• Learning Rate: 0.0001

Results:

After training the model for 100 epochs we got the following results:

1. train_loss:0.0921
2. val_loss: 0.0926
3. test_loss: 0.05484

Files in Order

• models.py
• dataset.py
• pl_modules.py
• augmentations.py
• configs.py
• utils.py
• Notebook Glue_tube_length_by_keypoint_detection.ipynb

DL Frameworks

• Pytorch
• Pytorch Lightning

Tensorboard monitoring

you can use tensorboard for monitoring the training. Use the following command after starting training.
tensorboard --logdir ./model/logs