How aparnamol-ks/vidsafe works
VidSafe is a video moderation application: a user submits a video, the UI runs the pipeline, and the system returns a selectively moderated video plus policy evidence and a report. The pipeline coordinates separate audio and visual analysis, merges the censored audio with the blurred video, aligns detections, and evaluates fused events against policies. The README documents multimodal detection and policy-based reporting; source excerpts show the principal calls. Some UI and reasoning details are only partially sampled, so uncertain links are omitted.
Main entry points: pipeline.py VidSafe pipeline · audio_pipeline.py Audio analysis pipeline · video_pipeline.py Visual analysis pipeline · app.py Streamlit interface
Components 24
User interface
- Streamlit interface
ui/app.py - PDF report export
ui/pdf_utils.py - Report summarizer
ui/policy_reducer.py
Moderation workflow
- Application entry
main.py - VidSafe pipeline
modules/pipeline.py - Multimodal alignment
modules/fusion/aligner.py - Audio-video merge
modules/audio/merger.py - Web video conversion
modules/video/format_converter.py - Runtime thresholds
config.py
Visual moderation
- Visual analysis pipeline
modules/video/video_pipeline.py - Semantic frame filter
modules/video/clip_filter.py - RT-DETR detector
modules/video/rtdetr_detector.py - Region blurring
modules/video/blur.py - Temporal consistency
modules/video/temporal_filter.py - Detection segments
modules/video/segment_builder.py
Audio moderation
- Audio analysis pipeline
modules/audio/audio_pipeline.py - Audio extraction
modules/audio/extractor.py - Speech transcription
modules/audio/transcriber.py - Word toxicity detection
modules/audio/roberta_detector.py - Sentence toxicity detection
modules/audio/detoxify_detector.py - Audio censoring
modules/audio/censor.py
Policy reasoning
- Policy reasoning
modules/reasoning/rag_engine.py - Moderation policies
modules/reasoning/policies/youtube_multimodal_policies.json
Other components
- Video submitter
Connections 26
- Video submitter → Streamlit interface: submits video
README.md - Streamlit interface → VidSafe pipeline: runs analysis
ui/app.py - Application entry → VidSafe pipeline: invokes
main.py - VidSafe pipeline → Audio analysis pipeline: runs audio analysis
modules/pipeline.py - VidSafe pipeline → Visual analysis pipeline: runs visual analysis
modules/pipeline.py - VidSafe pipeline → Audio-video merge: merges moderated media
modules/pipeline.py - VidSafe pipeline → Web video conversion: converts output
modules/pipeline.py - VidSafe pipeline → Multimodal alignment: aligns detections
modules/pipeline.py - VidSafe pipeline → Policy reasoning: evaluates evidence
modules/pipeline.py - Visual analysis pipeline → Semantic frame filter: filters frames
modules/video/video_pipeline.py - Visual analysis pipeline → RT-DETR detector: detects regions
modules/video/video_pipeline.py - Visual analysis pipeline → Temporal consistency: filters detections
modules/video/video_pipeline.py - Visual analysis pipeline → Detection segments: builds segments
modules/video/video_pipeline.py - RT-DETR detector → Region blurring: blurs regions
modules/video/rtdetr_detector.py - Visual analysis pipeline → Runtime thresholds: reads threshold
modules/video/video_pipeline.py - Audio analysis pipeline → Audio extraction: extracts audio
modules/audio/audio_pipeline.py - Audio analysis pipeline → Speech transcription: transcribes speech
modules/audio/audio_pipeline.py - Audio analysis pipeline → Word toxicity detection: scores toxic words
modules/audio/audio_pipeline.py - Audio analysis pipeline → Sentence toxicity detection: scores sentences
modules/audio/audio_pipeline.py - Audio analysis pipeline → Audio censoring: censors toxic spans
modules/audio/audio_pipeline.py - Speech transcription → Audio extraction: uses audio utilities
modules/audio/transcriber.py - Audio analysis pipeline → Runtime thresholds: reads censor settings
modules/audio/audio_pipeline.py - Policy reasoning → Moderation policies: loads policy data
modules/reasoning/rag_engine.py - Streamlit interface → PDF report export: generates PDF
ui/app.py - Streamlit interface → Report summarizer: uses report summarizer
ui/app.py - Report summarizer → Video submitter: requests report generation