Independent from-scratch reproduction of NVIDIA CVPR 2025 RL-RC-DoT: a block-level RL agent (PPO + Stable-Baselines3) for task-aware video compression. The agent outputs macroblock-level delta-QP maps injected into x264 to optimize downstream car detection (YOLO) on BDD100K under a bit-rate constraint.
reinforcement-learning object-detection reproduction x264 rate-control video-compression ppo bdd100k yolov5 stable-baselines3 cvpr-2025 task-aware-compression
-
Updated
Aug 3, 2026 - Python