Improving Single Shot Object Detection With Feature Scale Unmixing.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Single-shot detectors offer real-time performance but struggle with object scale variations.
• Current methods use multi-pyramid layers for scale-aware predictions, yet features lack sufficient scale awareness.
• Scale variations lead to missed small objects (false negatives) and partial detections of large objects (false positives).

Purpose of the Study:

• To propose a novel mechanism for feature scale-unmixing to improve scale-aware object detection.
• To address limitations in single-shot detectors caused by object scale variations.
• To develop a more accurate and efficient single-shot object detection network.

Main Methods:

• Introduced a Neighbor Erasing and Transferring (NET) mechanism for feature scale-unmixing.
• Designed a Neighbor Erasing Module (NEM) to emphasize small object features in shallow layers.
• Developed a Neighbor Transferring Module (NTM) to highlight large object features in deep layers.
• Constructed a single-shot network, NETNet, incorporating the NET mechanism and nearest neighboring pyramid feature aggregation.

Main Results:

• NETNet demonstrated effectiveness on the MS COCO and UAVDT datasets.
• Achieved 38.5% AP at 27 FPS and 32.0% AP at 55 FPS on the MS COCO dataset.
• Showcased improved performance in handling object scale variations.

Conclusions:

• The proposed NET mechanism effectively enhances scale-aware feature extraction in single-shot detectors.
• NETNet achieves a superior trade-off between real-time performance and detection accuracy.
• The method successfully mitigates false negatives and false positives related to object scale.