Abstract: Advanced object detectors often achieve high performance by using multi-scale feature fusion mechanisms. Most models utilize feature pyramids, either top-down or bottom-up path aggregation, and novel asymptotic fusion structures. However, the classical feature pyramid may cause semantic gaps in non-adjacent features, and the blind direction of asymptotic fusion may result in contamination or prejudice of lower-level feature information. This paper analyzes the potential defects of existing methods and designs the Top-down Asymptotic Feature Pyramid Network to address them. The proposed method is designed with two core components: Top-down Asymptotic Architecture and Adaptive Spatial Feature Fusion with Squeeze Weights, which respectively reduce the influence of higher-level features on lower-level feature information during feature fusion and enhance the adaptability of spatial feature integration by ensuring balanced weighting at the fusion stage. Furthermore, this paper also presents a more lightweight network by incorporating Depthwise Separated Convolution. We fairly compare the proposed method with various recent feature fusion approaches, applying it to both two-stage and one-stage object detectors, and evaluate their performance on authoritative datasets. Finally, we implement ablation experiments to verify and analyze the validity of the designed methods. Experiments show that our method obtains more accurate results than other modern feature pyramid networks. Codes are available on https://github.com/Iwill-github/TAFPN.
Keywords: Object detection, Feature pyramid network, Top-down asymptotic fusion, Adaptive spatial feature fusion
The code will be open source soon.