Neural Attention-Driven Non-Maximum Suppression for Person Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Non-maximum suppression (NMS) is crucial for object detection, refining candidate regions (RoIs).
• Standard GreedyNMS struggles with occlusions and requires manual tuning, impacting person detection accuracy.
• Existing methods lack robustness in complex scenarios like crowded scenes.

Purpose of the Study:

• To propose an efficient deep neural architecture for NMS tailored to person detection.
• To address the limitations of GreedyNMS in handling occlusions and overlapping detections.
• To achieve precise one-to-one detection assignment for each person instance.

Main Methods:

• Introduced Seq2Seq-NMS, a novel deep neural architecture for NMS.
• Formulated NMS as a sequence-to-sequence problem, leveraging Multihead Scale-Dot Product Attention.
• Jointly processed geometric and visual features of candidate RoIs.

Main Results:

• Seq2Seq-NMS demonstrated favorable results on three public person detection datasets.
• The proposed method outperformed competing NMS techniques.
• Achieved acceptable inference runtime, suitable for practical applications.

Conclusions:

• Seq2Seq-NMS offers an effective solution for accurate person detection, particularly in challenging occluded scenarios.
• The sequence-to-sequence formulation and attention mechanism enhance NMS performance.
• This architecture provides a robust and efficient alternative to traditional NMS algorithms.