bb_pipeline_models

This repository contains models used for running detection, saliency (localizer), and tracking in the bb_pipeline. Models are provided for compatibility with both Keras 2 and Keras 3.

Model Files and Compatibility

The following is a summary of the models, their purpose, and their compatibility with different versions of Keras.

Detection Step

Localizer Model (Saliency)

• Keras 3 Version: localizer_2019_keras3.h5 and localizer_2019_attributes.json
  • Additional attributes (localizer_2019_attributes.json) are required for using the model in Keras 3.
• Keras 2 Version: .model files (e.g., localizer_2019.model)

Decoder Model

• Keras 3 Version: decoder_2019_keras3.h5
• Keras 2 Version: decoder_2019.model

Tracking Step

• Model Files: detection_model_4.json, tracklet_model_8.json
  • These tracking model files in JSON format are compatible with both Keras 2 and Keras 3.

Folder Structure and Model Generation

models Folder

This folder contains the models used in the pipeline, organized by task:

• decoder: Contains models for the detection step.
• saliency: Contains models for the saliency (localizer) step.
• tracking: Contains tracking models compatible with both Keras 2 and Keras 3.

model_generation Folder

This folder provides notebooks and code for model conversion and generation:

• keras2_model_conversion.ipynb: Notebook to convert existing Keras 2 models to a format compatible with Keras 3 by saving the model weights.
• keras3_model_generation.ipynb: Notebook with functions to create model structures compatible with Keras 3 by importing model weights.
• pipelinemodels.py: Contains the model structure functions, which generate models for the decoder and localizer steps.