[论文解读] Learning Object Detectors from Scratch with Gated Recurrent Feature Pyramids.

Zhiqiang Shen, Humphrey Shi+7 more|arXiv (Cornell University)|Dec 4, 2017

Advanced Neural Network Applications参考文献 33被引用 37

一句话总结

本文提出了一种门控循环特征金字塔网络，通过在不同特征尺度上动态调整监督，从零开始学习目标检测器。通过将参数量减少至DSOD的1/3，并采用门控控制的特征精炼机制，该模型在PASCAL VOC 2012（VOC 07++12）上实现了77%的mAP，优于以往从零训练的方法，甚至超越了部分ImageNet预训练模型。

ABSTRACT

In this paper, we propose gated recurrent feature pyramid for the problem of learning object detection from scratch. Our approach is motivated by the recent work of deeply supervised object detector (DSOD), but explores new network architecture that dynamically adjusts the supervision intensities of intermediate layers for various scales in object detection. The benefits of the proposed method are two-fold: First, we propose a recurrent feature-pyramid structure to squeeze rich spatial and semantic features into a single prediction layer that further reduces the number of parameters to learn (DSOD need learn 1/2, but our method need only 1/3). Thus our new model is more fit for learning from scratch, and can converge faster than DSOD (using only 50% of iterations). Second, we introduce a novel gate-controlled prediction strategy to adaptively enhance or attenuate supervision at different scales based on the input object size. As a result, our model is more suitable for detecting small objects. To the best of our knowledge, our study is the best performed model of learning object detection from scratch. Our method in the PASCAL VOC 2012 comp3 leaderboard (which compares object detectors that are trained only with PASCAL VOC data) demonstrates a significant performance jump, from previous 64% to our 77% (VOC 07++12) and 72.5% (VOC 12). We also evaluate the performance of our method on PASCAL VOC 2007, 2012 and MS COCO datasets, and find that the accuracy of our learning from scratch method can even beat a lot of the state-of-the-art detection methods which use pre-trained models from ImageNet. Code is available at: this https URL .

研究动机与目标

• 为解决从零训练高精度目标检测器的挑战，特别是小目标检测问题。
• 减少特征金字塔网络中可学习参数的数量，以提升训练效率与收敛速度。
• 开发一种动态监督机制，根据目标尺度在训练过程中自适应调整。
• 通过在不同特征层级选择性地增强或抑制监督，提升小目标检测性能。
• 在PASCAL VOC和MS COCO数据集上实现最先进性能，且不依赖ImageNet预训练。

提出的方法

• 引入一种循环特征金字塔结构，通过多尺度迭代精炼特征，减少可学习参数数量。
• 网络采用门控机制，根据输入目标大小自适应调整不同特征层级的监督强度。
• 门控机制根据检测目标的尺度增强或抑制特征图，从而提升小目标检测能力。
• 整个架构从零开始端到端训练，无需依赖ImageNet预训练。
• 该方法将可学习参数数量减少至DSOD的1/3，仅需50%的迭代次数即可实现更快收敛。
• 通过融合多尺度的空间与语义信息，利用循环精炼过程更新特征金字塔。

实验结果

研究问题

• RQ1 循环特征金字塔架构能否提升目标检测从零训练的效率与性能？
• RQ2 通过门控机制实现的自适应监督对检测精度（尤其是小目标）有何影响？
• RQ3 能否从零训练的模型超越依赖ImageNet预训练的最先进检测器？
• RQ4 减少可学习参数数量在多大程度上提升了收敛速度与模型效率？
• RQ5 所提方法在PASCAL VOC与MS COCO等多样化数据集上的泛化能力如何？

主要发现

• 所提方法在PASCAL VOC 2012（VOC 07++12）上达到77% mAP，显著优于此前从零训练的SOTA方法（64%）。
• 在仅PASCAL VOC 2012数据集上，该方法达到72.5% mAP，证明其在无任何预训练情况下仍具备强大性能。
• 模型仅需DSOD所需迭代次数的50%即可收敛，表明因参数量减少而实现更快训练。
• 该方法在PASCAL VOC 2007与2012上均超越了众多使用ImageNet预训练的SOTA检测器。
• 门控控制的监督机制通过动态调节特征学习强度，显著提升了小目标检测性能。
• 在MS COCO上的表现证实了模型的泛化能力，无需预训练即可达到具有竞争力的准确率。