ENERGY-EFFICIENT VIDEO OBJECT DETECTION
A technique to increase efficiency in video object detection.
Video Object DetectionVideo ProcessingDeep Reinforcement Learning
Overview
This project aims at reducing the latency of computation in video object detection. The technique takes advantage of the redundancy of information commonly present in image sequences. Also, the strategy mimics the behavior of the human vision, which dynamically ignores the object context when performing a focused tracking. A novel technique in reinforcement learning was applied, where an agent learns over a distribution of reward funcionts, and can be conditioned at inference at on one particular function.
Technologies & Skills
Deep Reinforcement LearningVideo Object DetectionCPU Optimization
Key Achievements
- ✓Average latency reduction up to 2.09×
- ✓FPS rates in CPU similar to GPU
- ✓Novel deep reinforcement learning technique with reward-conditional training
- ✓Code publicly released
- ✓Open-source dataset released
Video
Images
🔍