PDNet: Towards Better One-stage Object Detection with Prediction Decoupling.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current one-stage object detectors use a per-pixel prediction approach, inferring both object category and boundary from each grid location.
• This shared inference location can lead to suboptimal performance as optimal positions for classification and localization often differ.

Purpose of the Study:

• To analyze optimal inference positions for object category and boundaries in object detection.
• To propose a novel prediction-target-decoupled detector (PDNet) for a more flexible detection paradigm.

Main Methods:

• PDNet employs a prediction decoupling mechanism to encode different targets separately in distinct locations.
• A learnable prediction collection module utilizes dynamic boundary and semantic points to aggregate predictions from favorable regions.
• A two-step strategy learns dynamic point positions by first estimating prior positions and then predicting residual offsets.

Main Results:

• Experiments on the MS COCO benchmark demonstrate PDNet's effectiveness and efficiency.
• Using a ResNeXt-64x4d-101-DCN backbone, PDNet achieved 50.1 AP with single-scale testing, surpassing state-of-the-art methods.
• PDNet maintains high efficiency as a one-stage framework.

Conclusions:

• PDNet offers a more flexible and effective object detection paradigm by decoupling prediction targets.
• The proposed method achieves state-of-the-art performance while maintaining the efficiency of one-stage detectors.
• The study highlights the benefits of tailored inference locations for different object detection targets.