Week 05

Jan 30

• we want to run the DNV with a deeper resnet.

  • timm library https://huggingface.co/docs/timm/index
  • https://walkwithfastai.com/vision.external.timm
  • ResNetV2 and BeIT models

• different restnet50, restnet101 -> test the images DNV and SD1

  • vertical layout
  • validation vs test score

• restnet101 doing on DNV

• DNV model fine tune with SD1 doing on SD1 data

Jan 31

• copy the email sent to sean regarding COF in my outline

• sqlite

• postgis

• do point in polygon function... for FME

• submit the resnet50 result on portal (need to email Joakim to get name on the portal)

• run job for SewerML resnet101

• azure - run for 20 images or so to see what kind of bounding box they dectect

• UBC azure credit - ask IT people

• text recognition

  • detect text bounding box
  • extract the text from the bounding box

Feb 01

• run azure on images of different annotation software

• get the count of images base on access and video

  • how many images getting added when doing continuous defect

• doing the dataflow diagram (vertical) from labelling image to model training to result

  • doing summary dataflow diagram (example we using this app to view the result and stuff)
  • doing sub dataflow diagram for labelling image

• continuous defect:

  • doing correlation images between the continous defect (between the start image or end image)
  • different camera orientation

• Ask PACP instructor about the old PACP codes

• Ask Sean if DNV prefer using WRc or PACP.

  • Sean replied using PACP

• How we deal with the historical data that has the different name.

  • Junction (old term)
  • new term need to specify what type of junction
  • Surface Damage need to modifier

• Talk about how we use SewerML model and got good results but did not use for North America utility because they are using different defect code

• create bounding box from DEM to get polygon

• move the collimate note from meeting with brian to report issue

Feb 02

• look at the bug in Grafana
• Easy OCR: (issue in dnv streamlit)
  • we might be able to get better result if we apply image augmentation
  • Sudhir suggest convert the pixel that completely black to white and other color as black
• Presentation for sean
  • define F1 and F2 score as true positive and true negative. F1 is defined first
  • then accurancy
  • best result so far
  • Having example for images that got define correctly from streamlit

Feb 03

• run with resnet152

• run SD1 model with DNV (video annotation) resnet101

• change pos_weight to weight for `loss=nn.BCEWithLogistsLoss(weigth=tw)

• take out all runs with image argumentation and reorganize

• defect vs no defect DNV. (just my idea: tell where location of defect are)

• future research: detect the distance from video. using spatial correlation to see if to know if video is moving forward

• send Dr.Lence reminding her about chris woo's email

• three screen shots from SD1 (30 test images)