A Reinforcement Learning-Based Automatic Video Editing Method Using the Knowledge from Vision-Language Model
Published in October 27, 2023
Research assistant
Engineered a distributed Python data-collection service (WebSockets + RabbitMQ) that streamed ≈ 18k annotated editing actions from 40 volunteers to a 4-GPU EC2 cluster in under 10 hours.
Containerized training and inference with Docker/TorchServe and shipped them as micro-services on an Ubuntu cloud node; a REST/gRPC backend and a WebSocket client scale independently, supporting rolling updates.
Authored an RTSP-decoder micro-service that resolves camera IPs, decodes multi-channel streams, and exposes a control API so the server can remotely open client video windows – maintaining ≤ 300 ms end-to-end latency.