07-May-2023

• Added r3d_18_inference.py to perform inference using Inception 3D model on provided video file
• Added load_video_from_dataset.py to obtain a randomly provided video from Data_Loader.py

To do:

• Extract feature from interim layer
• Read up the Inception architecture

08-May-2023

• Coded the feature extractor layer

To do:

• Build a simple classification model based on the feature extractor of a few video clips.
• Obtain anomaly detection accuracy.
• Build a UMAP dimensionality reducer for different anomaly types
• Compare current feature extractor to embeddings from a vision transformer

20-May-2023

• Built a CIFAR 10 classifier with custom images to be used for to process the interim layer of video feature extractor

24-Jun

• Video resnet model uses 3D convolutions with variable length output in the penultimate layer
• This doesn't allow the penultimate Layer 4 to be used for downstream learning tasks that need a fixed length feature input
• used the torch summary library to review output model sizes based on given video input - very useful
• Reverted to last layer of global average pool layer which returns (512) size output
• Questions
  • Will attention be needed given that temporal features are detected in the Video res net model (R3D)