Overview

This repository contains a Python-based tool for analyzing and processing photography model predictions, specifically focusing on frame analysis in video content.

Features

• Multiple Dataset Support:
  • Supports three main datasets:
    • ACT75 (75 videos dataset)
    • BF50 (Benchmark dataset)
    • FLASH (ActivityNet Captions dataset with 1k+ videos)
• Model Analysis Tools:
  • Frame prediction analysis
  • Logit data processing
  • Prediction synthesis across different window sizes
  • Model performance evaluation

Repository Structure

photography-model/
├── predictions/       # Model predictions and configuration
├── logits/            # Raw Logits
├── plots/             # Plots of the outputs
├── data/              # The BF50 and ACT75 datasets are stored here
├── results.ipynb      # Main results and analysis notebook
├── utils.py           # Core utility functions
├── benchmark.py       # Code & Functions for benchmarking
└── temp.ipynb         # Development and testing notebook

Core Features

1. Data Loading and Processing:

   • Support for JSON and pickle data formats
   • Ground truth data management for model evaluation

2. Prediction Synthesis:

   • Window size adjustments for predictions
   • Smoothed prediction calculations

3. Model Evaluation:

   • Support for multiple model comparisons
   • Frame-by-frame analysis
   • Performance metrics calculation

Getting Started

1. Ensure you have Python 3.x installed with Jupyter support
2. Install required dependencies
3. Clone this repository
4. Open results.ipynb to view analysis results or run your own experiments

Contributing

Feel free to submit issues and enhancement requests. Please ensure any pull requests maintain the project's code structure and documentation standards.