深度增強人群計數

Abstract

人群計數是一項用來計算在圖片中出現總人數的技術，這是一個經典且相當關鍵的問題，也被廣泛使用在不同的應用當中。目前的方法大多使用圖片的RGB值來預測，同時也達到不錯的效果。然而，當前的方法在預測擁擠區域時會受到不完整或是模糊的邊界影響。因此在論文中，我們提出了一個深度增強人群計數網路，這個網路使用了我們所提出的一項新技術，雙向跨型態專注機制，來融合深度的資訊。使用深度的資訊，可以讓我們的方法根據深度的值來專注在處理相對擁擠的區域。我們所提出的雙向跨型態專注機制，會藉由各自學習含有較多資訊量的區域，來交互式的融合不同的輸入型態。在我們的實驗中，我們的方法在目前最大且最難的資料集中打敗了目前最佳的模型。視覺化的結果說明了使用深度的資訊可以精準的預測擁擠的區域。最後，在敏感度分析中也證明了我們提出的每個元件都對最終的結果有正面的影響。

Crowd counting which aims to calculate the number of total instances on an image is a classic but crucial task that supports many applications. Most of the prior works are based on the RGB channels on the images and achieve satisfied performance. However, previous approaches suffer from counting highly congested region due to the incomplete and blurry shapes. In this paper, we present an effective crowd counting method, Depth Enhanced Crowd Counting Network (DECCNet), which leverages the estimated depth information with our novel Bidirectional Cross-modal Attention (BCA) mechanism. Utilizing the depth information enables our model to explicitly learn to pay attention to those congested regions on the basis of the depth information. Our BCA mechanism interactively fuses two different input modalities by learning to focus on the informative parts according to each other. In our experiments, we demonstrate that DECCNet outperforms the state-of-the-art on the two largest crowd counting datasets available, including UCF-QNRF, which has the highest crowd density. The visualized result shows that our method can accurately regress dense regions through leveraging depth information. Ablation studies also indicate that each component of our method is beneficial to final prediction.