Untemporal

Extend what you can see in a crime video through 3D representation Team code: 357CF5B72C78BB2F

Comment

Inspiration

In an era of deepfakes and video manipulation, verifying video authenticity is critical. We were inspired by the need for accessible forensic tools that can detect temporal inconsistencies—like cuts, reordering, and frame manipulation—without requiring expert knowledge. Whether it's evidence in legal cases, news verification, or social media content, people need to trust what they see.

What it does

Untemporal is a web-based video forensics tool that automatically detects tampering in video footage. Users upload a video, and the system analyzes it for temporal inconsistencies using 3D scene reconstruction and computer vision techniques. The results are visualized through an interactive timeline showing "spikes" where edits were detected, along with visual artifacts highlighting the inconsistencies. The system can detect:

  • Frame cuts - Missing segments that create visual jumps
  • Reordering attacks - Scenes rearranged to violate causality
  • Temporal inconsistencies - Objects appearing/disappearing unnaturally

How we built it

Backend (Python/FastAPI):

  • Frame extraction and preprocessing pipeline
  • 3D scene reconstruction using Structure-from-Motion (SfM) and Gaussian Splatting
  • Temporal consistency analysis comparing reconstructed 3D models across frames
  • SSIM (Structural Similarity Index) and L2 error metrics
  • Spike detection algorithm to identify anomaly frames

Frontend (React):

  • Video upload interface
  • Interactive timeline chart with visual spike indicators
  • 3D Gaussian Splat viewer for scene visualization
  • Before/after/diff view for detected anomalies
  • Real-time verdict display ("Likely edit detected at...")

Infrastructure:

  • Docker containerization for easy deployment
  • RESTful API contract between frontend and backend
  • Mock data generator for development and testing

Challenges we ran into

  1. Large file handling - 180MB+ Gaussian splat files exceeded GitHub's limits, requiring careful git management
  2. Performance optimization - 3D reconstruction is computationally expensive; we had to balance accuracy with demo speed
  3. Temporal analysis complexity - Detecting causality violations (like objects appearing before they should exist) required careful algorithm design
  4. Integration complexity - Coordinating three independent components (forensics logic, backend pipeline, frontend) with minimal dependencies
  5. Real-time processing - Making the analysis fast enough for a live demo while maintaining accuracy

Accomplishments that we're proud of

  • Built a complete end-to-end system from video upload to forensic analysis in a hackathon timeframe
  • Successfully detected both cut-based and reordering-based tampering attacks
  • Created an intuitive UI that makes complex forensic analysis accessible to non-experts
  • Achieved sub-2-minute processing time per video clip for demo purposes
  • Implemented a modular architecture that allows backend improvements without frontend changes
  • Successfully visualized 3D scene inconsistencies in an interactive web interface

What we learned

  • The complexity of video forensics and the importance of temporal consistency
  • How 3D scene reconstruction can reveal 2D video manipulation
  • The challenges of working with large 3D model files in version control
  • The value of clear API contracts when building distributed systems
  • How to balance technical accuracy with user experience in specialized tools
  • The importance of fallback strategies when GPU-intensive operations aren't available

What's next for Untemporal

  • Real-time processing - Optimize the pipeline to handle longer videos in real-time
  • Advanced detection - Add support for detecting deepfake-style manipulations
  • Batch processing - Allow users to analyze multiple videos at once
  • Export capabilities - Generate forensic reports with detailed evidence
  • API expansion - Open the API for integration with other forensic tools
  • Machine learning enhancement - Train models to improve detection accuracy
  • Mobile support - Develop a mobile app for on-the-go video verification
  • Cloud deployment - Scale the service to handle production workloads
Share this project:

Updates