🤖 AI Summary
In multi-scale object detection, point-wise fusion in conventional feature pyramids causes misalignment between features across pyramid levels. Method: This paper proposes the Independent Hierarchical Pyramid (IHP) architecture, abandoning traditional top-down/bottom-up fusion paradigms. It introduces Soft Nearest-Neighbor Interpolation (SNI) and Extended Spatially Adaptive Downsampling (ESD) to establish an end-to-end secondary alignment (SA) mechanism—enabling precise, lightweight cross-scale feature matching without significant computational overhead. The method leverages GSConvE, a lightweight convolutional design, to enhance feature consistency without additional parameters. Contribution/Results: Evaluated on Pascal VOC and MS COCO, IHP achieves real-time state-of-the-art detection performance, with substantial AP gains for small objects. It jointly optimizes inference speed and accuracy, demonstrating superior efficiency–accuracy trade-offs.
📝 Abstract
Multi-head detectors typically employ a features-fused-pyramid-neck for multi-scale detection and are widely adopted in the industry. However, this approach faces feature misalignment when representations from different hierarchical levels of the feature pyramid are forcibly fused point-to-point. To address this issue, we designed an independent hierarchy pyramid (IHP) architecture to evaluate the effectiveness of the features-unfused-pyramid-neck for multi-head detectors. Subsequently, we introduced soft nearest neighbor interpolation (SNI) with a weight downscaling factor to mitigate the impact of feature fusion at different hierarchies while preserving key textures. Furthermore, we present a features adaptive selection method for down sampling in extended spatial windows (ESD) to retain spatial features and enhance lightweight convolutional techniques (GSConvE). These advancements culminate in our secondary features alignment solution (SA) for real-time detection, achieving state-of-the-art results on Pascal VOC and MS COCO. Code will be released at https://github.com/AlanLi1997/rethinking-fpn. This paper has been accepted by ECCV2024 and published on Springer Nature.