List of Tasks

Literature Review

• Thesis outline
• What are the practice during sewer inspection?
• Technology used in Sewer inspection
• PACP standard vs Danish standard
• Why we use fastai over other algorithm for multi-label classification
• The benefit of using rest50 as a backbone and doing three fine turning process (resnet50->Sewer-ML->DNV data). Is there any benefit from it
• Different metrics used to evaluate the model's performance (such as F1 score vs F2 score vs Accuracy)
• Different batch size and image size affect GPU and computation time
• Different way to photoshop text (Gaussian blur, Pixel by Pixel (interpolation), Find certain pixel value)
• Read Joakin's dissertation for more information regarding Danish standard.

Multi-label classification

Sewer-ML

• Train fastai on Sewer-ML training image
• Optimal learning rate
• Optimal image size (currently trying resize(224, squish) and resize(712,478) (DNV image size))
• Optimal batch size
• Utilize all four node of GPUs. Launching_distributed_code
• Hyper parameter tuning the model using sweep
• Resume_Training_From_Checkpoint
• Fine_tune_vs_fit_one_cycle
• Add F2 score and F1 score as evaluation metrics.
• Use Sewer-ML validation image to get the metric (such F1-normal, F2 score CWI)
• Need to find out what the model do to testing image when doing prediction (auto pooling, resize?)
• Use Sewer-ML to predict the label of DNV image
• Need to find out how they mark certain image as increasing in 10% waterflow?
• Read about Danish_inspection_section2.
• Read about why the four levels of VA water_level_in_Danish_inspection_link1.
• water_level_in_Danish_inspection_link2

DNV dataset

• Create a label csv files for each image (using whisper to grab the text in the image and verify whether we can use this information to classify each image with label and without label) whisper
• DNV and Sewer Label have different label status. Create .yml file to replace DNV label with Sewer-ML label standard when we use Sewer-ML to predict label of DNV image
• Train model with DNV data without backbone (such as resnet50). Hyper parameter tuning the model
• Train model with three stages fine turning (resnet50->Sewer-ML-> DNV). Hyper parameter tuning the model
• Train model with image with text vs images without text (photoshop version)

Regression

• Find out if regression help with image segmentation and object detection

Object Detection

• Using YOLOv7 to do object detection

Image Segmentation

• Tool use to annotate the image

Pipe Deterioration Model

• Get the report of the condition of the pipe after the sewer inspection
• Gather features of the sewer mains
• Develop the deterioration model using Machine learning using algorithms such as (RSF and XGBSE)
• Predict the pipe conditions of uninspected pipes