影像物件偵測後處理方法比較

Abstract

電腦視覺技術發展至今已能處理諸多領域的應用問題，從自動駕駛、醫療影像判讀到交通流量計算。近年來以深度學習方法為熱門研究對象，在諸多應用問題如影像分類、物件偵測、語義分割等方面都較傳統方法表現優異，而深度學習方法的物件偵測方法多以兩階段偵測器R-CNN架構延伸，得益於人工先驗方法如錨框設計、後處理方法、標記指定，但也受限於參數調整和密集預測的處理。

本文欲討論深度學習物件偵測的後處理方法對預測表現的影響，以貪婪非極大值抑制(Greedy NMS)方法作為基準方法，與Soft NMS、Fast NMS、Cluster NMS、DIOU NMS、Confluence NMS、Weight NMS等方法相比在公開資料集 PASCAL VOC、MS COCO上的表現差異與優缺點，並從不同模型與資料集的表現、類別與幾何因素的錯誤數、參數敏感度、執行速度切入。發現不同指標的表現主要受模型與資料集影響，後處理方法間的主要差異在於偽陰性錯誤數和偽陽性錯誤數的權衡以及對分類機率門檻值、定位門檻值的敏感度，速度上的差異也受參數影響，而與演算法關係不大。

Computer vision has become a common technique in a variety of applications including autonomous vehicles, medical image recognition, and traffic flow monitoring. In recent years, deep learning is the most popular study area in this domain which is capable of solving multiple problems including image classification, object detection, semantic segmentation, etc. Studies have shown that they perform much better than the conventional methods. Most of object detection models are inspired from the architecture of two stage detector R-CNN, and make improvements with artificial prior methods e.g. anchor design, post processing methods, label assignment. However, these prior methods subject to parameter tuning and processing methods for dense predictions.

In this paper, we aim to investigate the effect of post processing methods to the performance of deep learning object detectors and discuss their differences and pros/cons from macro aspect of dataset/model combinations to micro aspect of categories and geometric factors and parameter sensitivity and speed. With Greedy NMS as the baseline method, we compare it to Soft NMS, Fast NMS, Cluster NMS, DIOU NMS, Confluence NMS, Weight NMS, with their performance on PASCAL VOC, MS COCO datasets. The results reveal that the performance of different metrics is primarily influenced by the models and datasets. The primary difference between different post-processing methods is how each method is capable of balancing false negative rate and false positive rate, as well as the sensitivity to classification probability thresholds and localization thresholds. In addition, time complexity depends only on the parameters and has little effect from the algorithms employed.