Skip to main content

List of Findings

DNV data​

  • The DNV video_lists can be found in the /media/gqc/unionsine1/VS_Research/CCTV/DNV/Data/Video_Lists directory. The CSV files in this directory contain the name of the video along with its video_type.

09/05/2024​

Definition of terms used in CCTV CV:​

  1. Dataset: A Dataset is composed of a set of Video Groups.
  2. Video Group: A Video Group composed of a set of inspection videos.
  3. frames: Images extracted from an inspection video.

Usually a Dataset is named as <utility_name>_<number_of_videos_received>. The Video Group name can be arbitrary but it has some context clues such as the date when the inspection videos that belong to that Video Group were received or which videos are present in that Video Group.

Predictions made by SD1_unblurred_VB_model on blurred frames:​

Identified 24,881 frames where defect is not ND and threshold > 0.2. 43 distinct defects were identified by this model.

09/06/2024​

  • Used Video_DB_SD1_sample_01202023.db to create training and testing CSVs.
  • Unzipped the blurred images in the 01202023 video_group and looked at the images themselves. The images looked fine.
  • Trained a Frame Based (FB) model using the 01202023 video_group.
  • wandb: accuracy_multi 0.99513 wandb: epoch 10 wandb: eps_0 1e-05 wandb: eps_1 1e-05 wandb: eps_2 1e-05 wandb: learning rate 0.001 wandb: lr_0 0.0 wandb: lr_1 0.0 wandb: lr_2 0.0 wandb: mom_0 0.9 wandb: mom_1 0.9 wandb: mom_2 0.9 wandb: raw_loss 0.03753 wandb: sqr_mom_0 0.99 wandb: sqr_mom_1 0.99 wandb: sqr_mom_2 0.99 wandb: train_loss 0.01794 wandb: train_samples_per_sec 43.58663 wandb: valid_loss 0.01405

The numbers were more reasonable. Maybe we need more epochs?