Meeting notes

October 28, 2022

DEVEN, VANNARY, PAVAN

Organization of data

1. Frame captured from the video: grouped by named of the video they came from
2. Response from the Google Vision API: grouped by named of the video they came from
3. CSV for labels: one for each video

Ideal solution would be to group them by labels.

Currently models are only being trained for 33 videos from Lynn Valey Oct 6. We have 40 more videos that have not been used for training.

Nuances discussed

1. There is a discrepancy between the frame number from all_frame csv and the frame number captured in the plot_spatial_correlation()

   PAVAN: After debugging we found that the issue was not in indexing. It was that the final image in the grid was a duplicate of immediately previous frame, which should have shown the current (selected) frame. It is fixed now.

2. The number of frames from where the defect is visible to the frame where the defect is labelled will vary. It depends on how long the inspector would pause to enter the observation and the amount of time they spend looking around the defect area.

3. The number of frames that we need to look back at before finding the frame at the different positions can be large.

Other nuances from the dataset

1. At the distance where there are multi labels, it takes about 30 sec for the investigator to enter between each label. Therefore, we will have about 30 frames of the same position between each label. Instead of looking back from the first frame of each label, we will need only to look back from the first image where the first label is annotated. If the image has a high correlation with the selected image, the image should be assigned to all the labels.