以高估圖優化語義分割中的異常物偵測

Abstract

在這篇論文中，我們提出了一個新方法，能夠用於語意分割(semantic segmentation)中，像素層面的異常偵測(out-of-distribution detection)。這個問題在像自動駕駛(autonomous driving)這樣的安全敏感領域是一個具有挑戰性的任務。最近在異常分割(anomaly segmentation)研究領域中的一個最佳方法是訓練網絡使用熵(entropy)來區分內部數據(inliers)和外部數據(outliers)。然而，這種方法經常導致許多假陽性。為了解決這個缺點，我們提出了一種高效的方法，結合了一個高估圖(overestimated map)和一個輕量級耦合模組(coupling module)。首先，我們使用分佈範圍外(out-of-distribution)的樣本重新訓練分割網絡，以放大這些分佈範圍外樣本的熵。然後，我們使用這個重新的訓練網絡產生的輸出來計算高估圖，並將輸出、熵圖和高估圖輸入耦合模組進行處理。最後通過將耦合模組的輸出與熵圖相乘獲得最終的異常圖(final anomaly map)。我們的方法在 SegmentMeIfYouCan 數據集上進行了評估，該數據集包括 AnomalyTrack、ObstacleTrack 和 LostAndFound 子數據集。與作為基準(baseline)的最佳方法相比，只增加了最小的推理時間開銷，便能降低最多 50% 的假陽性。

In this paper, we propose a new method for detecting anomalies at the pixel level in semantic segmentation. This is a challenging task in safety-sensitive domains, such as autonomous driving. A prominent method of anomaly segmentation involves training the segmentation network to differentiate inliers and outliers on the basis of entropy. However, this approach often leads to many false positives. To address this drawback, we propose an efficient method that involves combining an overestimated map with a lightweight coupling module. Initially, we retrain the segmentation network with an OoD proxy to amplify softmax entropy for these OoD samples. We then calculate the overestimated map using logits generated by this retrained network and input logits, entropy map, and overestimated map into the coupling module for processing. The final anomaly map is obtained by multiplying the output of coupling module with the entropy map. On the SegmentMeIfYouCan benchmark, including AnomalyTrack, ObstacleTrack, and LostAndFound, the proposed approach achieved up to 50% fewer false positives than did a state-of-the-art method (the maximized entropy method) at the cost of only a slightly longer inference time.