Conceitos essenciais
The author presents MONDRIAN, a system for high-performance object detection on high-resolution video streams using Compressive Packed Inference. The approach minimizes per-pixel processing costs and maximizes processing parallelism by dynamically scaling ROIs and packing them into large canvases.
Resumo
In the paper, MONDRIAN is introduced as an edge system for efficient object detection on high-resolution video streams. The system utilizes Compressive Packed Inference to optimize processing costs and maximize parallelism. Through various techniques, MONDRIAN outperforms existing baselines in accuracy and throughput.
The content discusses the challenges faced in high-resolution video analytics and the limitations of current lightweight models and system optimization techniques. It proposes a novel approach to address these challenges effectively. By dynamically rescaling ROIs and intelligently packing them into canvases, MONDRIAN achieves significant improvements in accuracy and throughput.
Key points include the introduction of MONDRIAN as an edge system for object detection, the innovative Compressive Packed Inference approach, challenges in safe area estimation and ROI packing, and additional optimizations for improved performance.
The evaluation showcases the superior performance of MONDRIAN compared to baseline systems across different datasets, models, and devices. Through detailed analysis and experimentation, the effectiveness of the proposed system is demonstrated.
Estatísticas
Evaluation shows MONDRIAN outperforms baselines by 15.0-19.7% higher accuracy.
×6.65 higher throughput than frame-wise inference for processing 1080p videos.
Safe areas can be scaled to only 21% of original area on average.
Proactive Scale Predictor predicts coarse-grained estimate of safe area per ROI.
Reactive Scale Tuner explores nearby candidate scales to obtain optimal safe area.