Train a CV model using DNV frames which have root based defects in it and use it on COV.
Figure out why DNV model performs worse than metro model.
Grouping up the different Streamlit Apps and giving them headers(Maybe djangofying the streamlit apps)
Maybe changing COV PNGs to JPGs.
For DNV, there are 1181 videos that we do not have, how do we know that they even exist?
07/15/24: Tasks that need to be done w.r.t model training
Look at the results of increasing the batch size (whether training is faster/slower).
Look into training the model with less number of ND images.
~~ Creating a Mermaid to show how the flag affects what and how many models get created. (Priority: Highest) ~~
Migrating to use DBs rather than CSVs during training. (Priority: Medium)
Loading the trained single stage model (cctv-multilabel-sd1_d_all_VB_larger.pth) in the streamlit carousel and checking the model's performance. (Priority: )
Training the two-stage models. (Priority: )
Documenting the flow of training the two-stage models (if not already present in Vannary's paper). (Priority: )
Training Frame Based models. (Priority: )
Comparing the trained Frame Based and Video Based models. (Priority: )
Documenting time taken to unzip images belonging to a SD1 video group. (Priority: )
Renaming the NO_ND flag to two_stage and changing all references to NO_ND to two_stage. (Priority: Low)
Redirecting the training process to pull images from one unzipped location instead of copying over to the train folder.