VIDEX is a semantic video indexing system that converts surveillance footage into searchable segments through integrated object detection, open-vocabulary expansion, and retrieval-enhanced scene captioning. The system enables rapid natural-language retrieval of incident-relevant scenes and reduces analytical workload through scalable distributed inference.
- Semantic video indexing
- Natural-language video retrieval
- Surveillance video
- Open-vocabulary detection
Demo video is avaiable in YouTube
The frontend of VIDEX offers user-friendly interface so that user can easily explore the result of analysis. Here are more detailed descroption about interface and how to use them.
- Create project by uploading video which want to analyze
- Set project name and crop the video in upload modal
- Sort the project
- Search the project names
- Delete created projects
- Video Player
- Video w/ Yolo&OWL-ViT Object Detection Bounding Box
- Play/Pause, Fast Forward, Volume Control, Full Screen
- Anomaly Score Graph
- Show LOF score computed by video segment embedding
- Search Panel
- Submit user query through Search Bar
- Retrieve the top-20 description and correcponding timestamp
- Combine consequtive similar description
- Object Confidence Graph
- Show multiple confidences of detected objects
- User can toggle each object class graph
- Object Detection Panel
- Show the result of object detection with thumbnail and correcponding timestamp
- Filter the result by toggling chips or select and setting confidence threshold