Is Heuristic Sampling Necessary in Training Deep Object Detectors?

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep object detection models struggle with extreme foreground-background imbalance.
• Heuristic sampling methods (hard and soft) are commonly used to address this imbalance, but can be complex.
• Previous research suggested these methods are essential for accurate training.

Purpose of the Study:

• To challenge the necessity of heuristic sampling methods in deep object detection.
• To propose a novel, simpler alternative for handling foreground-background imbalance.
• To improve the accuracy of object detectors without complex sampling strategies.

Main Methods:

• Proposed a Sampling-Free mechanism using initialization and loss scaling.
• Addressed unreasonable classification gradient magnitudes caused by imbalance.
• Validated the method on both anchor-based and anchor-free object detectors.

Main Results:

• The Sampling-Free mechanism achieved higher detection accuracy than heuristic sampling methods.
• Effectiveness demonstrated on COCO and PASCAL VOC datasets.
• The method requires no laborious hyperparameter searching.

Conclusions:

• Heuristic sampling methods are not strictly necessary for training accurate deep object detectors.
• The proposed Sampling-Free mechanism offers a simple, effective, and data-diagnostic approach.
• This work provides a new perspective for addressing foreground-background imbalance in object detection.