- 學位論文
改進單次多框偵測器架構與後處理
Improving Single Shot Multibox Detector with Feature Pyramid Networks Structure and Postprocessing for Better Object Detection Performance
摘要
近幾年， 電腦視覺(computer vision)領域大量地使用卷積類神經網路(CNNs)。在這份論文中，我們會以CNNs當作物體偵測(Object Detection)的基礎方法。我們會先從單次多框偵測器(Single Shot Multibox Detector, SSD)當作方法的基礎。我們在SSD上加上特徵金字塔網路(Feature Pyramid Networks, FPN)，讓每一個位置都有全域資訊與地域的資訊。我們也在後處裡(postprocessing)的時候，增加了圍框投票(bounding box voting)的方法，來獲得更好的定位效果。在實驗當中，我們主要使用的資料庫為Pascal VOC 2007 test。而在物體偵測的領域中，我們使用平均準確度(Average Precision, AP)來當作衡量的標準，我們會平均每個類別的AP，得到mean AP(mAP)。在原始的SSD中，我們可以得到77.21% mAP的結果 我們在論文的實驗中，比較我們改進過後的方法與原始的SSD以及其他的物體偵測架構。結果顯示，最終我們可以得到77.85% mAP的結果。我們的方法獲得了更好的偵測成果。
並列摘要
In recent years, Convolutional Neural Networks(CNNs) have gained a lot of popularity in computer vision. In this work, we will use convolutional neural networks for object detection. To start with, we use Single Shot Multibox Detector(SSD) [7] as our basic framework. We impose Feature Pyramid Networks on SSD to combine local and global information. We also adjust postprocessing with bounding box voting for better localization. For comparison, we test our model on Pascal VOC 2007 test dataset. During evaluation, we calculate Average Precision(AP) for each model and class. Then, we average each AP to get mean Average Precision(mAP) as our final evaluation metric. With original SSD, we can have 77.21% mAP in Pascal VOC 2007 test dataset. In this thesis, our simulation results show that, the proposed method outperforms the original SSD and has better performance for object detection. Our final model can achieve 77.85% mAP.